LEARNING STRUCTURED PREDICTION MODELS

:
A LARGE MARGIN APPROACH
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF COMPUTER SCIENCE
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Ben Taskar
December 2004
c _Copyright by Ben Taskar 2005
All Rights Reserved
ii
I certify that I have read this dissertation and that, in my opin-
ion, it is fully adequate in scope and quality as a dissertation
for the degree of Doctor of Philosophy.
Daphne Koller
Computer Science Department
Stanford University
(Principal Advisor)
I certify that I have read this dissertation and that, in my opin-
ion, it is fully adequate in scope and quality as a dissertation
for the degree of Doctor of Philosophy.
Andrew Y. Ng
Computer Science Department
Stanford University
I certify that I have read this dissertation and that, in my opin-
ion, it is fully adequate in scope and quality as a dissertation
for the degree of Doctor of Philosophy.
Fernando Pereira
Computer Science Department
University of Pennsylvania
Approved for the University Committee on Graduate Stud-
ies:
iii
iv
To my parents, Mark and Tsilya, and my love, Anat.
Abstract
Most questions require more than just true-false or multiple-choice answers. Yet super-
vised learning, like standardized testing, has placed the heaviest emphasis on complex
questions with simple answers. The acquired expertise must now be used to address tasks
that demand answers as complex as the questions. Such complex answers may consist of
multiple interrelated decisions that must be weighed against each other to arrive at a glob-
ally satisfactory and consistent solution to the question. In natural language processing, we
often need to construct a global, coherent analysis of a sentence, such as its corresponding
part-of-speech sequence, parse tree, or translation into another language. In computational
biology, we analyze genetic sequences to predict 3D structure of proteins, find global align-
ment of related DNA strings, and recognize functional portions of a genome. In computer
vision, we segment complex objects in cluttered scenes, reconstruct 3D shapes from stereo
and video, and track motion of articulated bodies.
We typically handle the exponential explosion of possible answers by building mod-
els that compactly capture the structural properties of the problem: sequential, grammat-
ical, chemical, temporal, spatial constraints and correlations. Such structured models in-
clude graphical models such as Markov networks (Markov random fields), recursive lan-
guage models such as context free grammars, combinatorial optimization problems such as
weighted matchings and graph-cuts. This thesis presents a discriminative estimation frame-
work for structured models based on the large margin principle underlying support vector
machines. Intuitively, the large-margin criterion provides an alternative to probabilistic,
likelihood-based estimation methods by concentrating directly on the robustness of the de-
cision boundary of a model. Our framework defines a suite of efficient learning algorithms
that rely on the expressive power of convex optimization to compactly capture inference or
vii
solution optimality in structured models. For some of these models, alternative estimation
methods are intractable.
The largest portion of the thesis is devoted to Markov networks, which are undirected
probabilistic graphical models widely used to efficiently represent and reason about joint
multivariate distributions. We use graph decomposition to derive an exact, compact, con-
vex formulation for large-margin estimation of Markov networks with sequence and other
low-treewidth structure. Seamless integration of kernels with graphical models allows ef-
ficient, accurate prediction in real-world tasks. We analyze the theoretical generalization
properties of max-margin estimation in Markov networks and derive a novel type of bound
on structured error. Using an efficient online-style algorithm that exploits inference in the
model and analytic updates, we solve very large estimation problems.
We define an important subclass of Markov networks, associative Markov networks
(AMNs), which captures positive correlations between variables and permits exact infer-
ence which scales up to tens of millions of nodes and edges. While likelihood-based meth-
ods are believed to be intractable for AMNs over binary variables, our framework allows
exact estimation of such networks of arbitrary connectivity and topology. We also intro-
duce relational Markov networks (RMNs), which compactly define templates for Markov
networks for domains with relational structure: objects, attributes, relations.
In addition to graphical models, our framework applies to a wide range of other models:
We exploit context free grammar structure to derive a compact max-margin formulation that
allows high-accuracy parsing in cubic time by using novel kinds of lexical information. We
use combinatorial properties of weighted matchings to develop an exact, efficient formu-
lation for learning to match and apply it to prediction of disulfide connectivity in proteins.
Finally, we derive a max-margin formulation for learning the scoring metric for clustering
from clustered training data, which tightly integrates metric learning with the clustering
algorithm, tuning one to the other in a joint optimization.
We describe experimental applications to a diverse range of tasks, including handwrit-
ing recognition, 3D terrain classification, disulfide connectivity prediction in proteins, hy-
pertext categorization, natural language parsing, email organization and image segmen-
tation. These empirical evaluations show significant improvements over state-of-the-art
methods and promise wide practical use for our framework.
viii
Acknowledgements
I am profoundly grateful to my advisor, Daphne Koller. Her tireless pursuit of excellence in
research, teaching, advising, and every other aspect of her academic work is truly inspira-
tional. I am indebted to Daphne for priceless and copious advice about selecting interesting
problems, making progress on difficult ones, pushing ideas to their full development, writ-
ing and presenting results in an engaging manner.
I would like to thank my thesis committee, Andrew Ng and Fernando Pereira, as well
as my defense committee members, Stephen Boyd and Sebastian Thrun, for their excellent
suggestions and thought-provoking questions. I have learned a great deal from their work
and their influence on this thesis is immense. In particular, Fernando’s work on Markov
networks has inspired my focus on this subject. Andrew’s research on clustering and clas-
sification has informed many of the practical problems addressed in the thesis. Sebastian
introduced me to the fascinating problems in 3D vision and robotics. Finally, Stephen’s
book on convex optimization is the primary source of many of the insights and derivations
in this thesis.
Daphne’s research group is an institution in itself. I am very lucky to have been a part
of it and shared the company, the ideas, the questions and the expertise of Pieter Abbeel,
Drago Anguelov, Alexis Battle, Luke Biewald, Xavier Boyen, Vassil Chatalbashev, Gal
Chechik, Lise Getoor, Carlos Guestrin, Uri Lerner, Uri Nodelman, Dirk Ormoneit, Ron
Parr, Eran Segal, Christian Shelton, Simon Tong, David Vickrey, Haidong Wang and Ming
Fai Wong.
Jean-Claude Latombe was my first research advisor when I was still an undergraduate
at Stanford. His thoughtful guidance and quick wit were a great welcome and inspiration
to continue my studies. When I started my research in machine learning with Daphne, I
ix
had the privilege and pleasure to work with Nir Friedman, Lise Getoor and Eran Segal.
They have taught me to remain steadfast when elegant theories meet the ragged edge of
real problems. Lise has been a great friend and my other big sister throughout my graduate
life. I could not wish for more caring advice and encouragement than she has given me all
these years.
I was lucky to meet Michael Collins, Michael Littman and David McAllester all in
one place, while I was a summer intern at AT&T. Collins’ work on learning large-margin
tagging and parsing models motivated me to look at such methods for other structured
models. I am very glad I had the chance to work with him on the max-margin parsing
project. Michael Littman, with his endearing combination of humor and warmth, lightness
and depth, has been a constant source of encouragement and inspiration. David’s wide-
ranging research on probability and logic, languages and generalization bounds, to mention
a few, enlightened much of my work.
(Almost) everything I know about natural language processing I learned from Chris
Manning and Dan Klein. Chris’ healthy scepticism and emphasis on the practical has kept
me from going down many dead-end paths. Dan has an incredible gift of making ideas
work, and better yet, explaining why other ideas do not work.
I would like to express my appreciation and gratitude to all my collaborators for their
excellent ideas, hard work and dedication: Pieter Abbeel, Drago Anguelov, Peter Bartlett,
Luke Biewald, Vassil Chatalbashev, Michael Collins, Nir Friedman, Audrey Gasch, Lise
Getoor, Carlos Guestrin, Geremy Heitz, Dan Klein, Daphne Koller, Chris Manning, David
McAllester, Eran Segal, David Vickrey and Ming Fai Wong.
Merci beaucoup to Pieter Abbeel, my sounding board for many ideas in convex opti-
mization and learning, for numerous enlightening discussions that we have had anytime,
anywhere: gym, tennis court, biking to EV, over cheap Thai food. Ogromnoye spasibo to
my officemate, sportsmate, always late, ever-ready to checkmate, Drago Anguelov, who
is always willing to share with me everything from research problems to music and back-
packing treks to Russian books and Belgian beer. Blagodarya vi mnogo for another great
Bulgarian I have the good luck to know, Vasco Chatalbashev, for his company and ideas,
good cheer and the great knack for always coming through. Muchas gracias to Carlos
Guestrin, my personal climbing instructor, fellow photography fanatic, avid grillmaster
x
and exploiter-of-structure extraordinaire, for opening his ear and heart to me when it mat-
ters most. Mh-goi-saai to Ming Fai Wong, for his sense of humor, kind heart and excellent
work.
Many thanks to my friends who have had nothing to do with work in this thesis, but
worked hard to keep my relative sanity throughout. I will not list all of you here, but my
gratitude to you is immense.
My parents, Mark and Tsilya, have given me unbending support and constant encour-
agement. I thank them for all the sacrifices they have made to ensure that their children
would have freedom and opportunities they never had in Soviet Union. To my sister, Anna,
my brother-in-law, Ilya, and my sweet niece, Talia, I am grateful for bringing me so much
joy and love. To my tireless partner in work and play, tears and laughter, hardship and love,
Anat Caspi, thank you for making me happy next to you.
xi
xii
Contents
Abstract vii
Acknowledgements ix
1 Introduction 1
1.1 Supervised learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Complex prediction problems . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Structured models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Previously published work . . . . . . . . . . . . . . . . . . . . . . . . . . 13
I Models and methods 15
2 Supervised learning 16
2.1 Classification with generalized linear models . . . . . . . . . . . . . . . . 19
2.2 Logistic regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Logistic dual and maximum entropy . . . . . . . . . . . . . . . . . . . . . 21
2.4 Support vector machines . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 SVM dual and kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3 Structured models 24
3.1 Probabilistic models: generative and conditional . . . . . . . . . . . . . . . 25
xiii
3.2 Prediction models: normalized and unnormalized . . . . . . . . . . . . . . 27
3.3 Markov networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.1 Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.2 Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.3 Linear programming MAP inference . . . . . . . . . . . . . . . . . 34
3.4 Context free grammars . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5 Combinatorial problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4 Structured maximum margin estimation 42
4.1 Max-margin estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.1 Min-max formulation . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.2 Certificate formulation . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Approximations: upper and lower bounds . . . . . . . . . . . . . . . . . . 50
4.2.1 Constraint generation . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2.2 Constraint strengthening . . . . . . . . . . . . . . . . . . . . . . . 52
4.3 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
II Markov networks 57
5 Markov networks 58
5.1 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . . . . . 59
5.2 Maximum margin estimation . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.3 M
3
N dual and kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.4 Untriangulated models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.5 Generalization bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.6 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6 M
3
N algorithms and experiments 75
6.1 Solving the M
3
N QP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.1.1 SMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
xiv
6.1.2 Selecting SMO pairs . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.1.3 Structured SMO . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.3 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7 Associative Markov networks 89
7.1 Associative networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.2 LP Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.3 Min-cut inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.3.1 Graph construction . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.3.2 Multi-class case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.4 Max-margin estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.6 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
8 Relational Markov networks 109
8.1 Relational classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.2 Relational Markov networks . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.3 Approximate inference and learning . . . . . . . . . . . . . . . . . . . . . 116
8.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.4.1 Flat models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.4.2 Link model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8.4.3 Cocite model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.5 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
8.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
III Broader applications: parsing, matching, clustering 127
9 Context free grammars 128
9.1 Context free grammar model . . . . . . . . . . . . . . . . . . . . . . . . . 128
xv
9.2 Context free parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.3 Discriminative parsing models . . . . . . . . . . . . . . . . . . . . . . . . 133
9.3.1 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . 134
9.3.2 Maximum margin estimation . . . . . . . . . . . . . . . . . . . . . 135
9.4 Structured SMO for CFGs . . . . . . . . . . . . . . . . . . . . . . . . . . 138
9.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
9.6 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
9.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
10 Matchings 144
10.1 Disulfide connectivity prediction . . . . . . . . . . . . . . . . . . . . . . . 145
10.2 Learning to match . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
10.3 Min-max formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
10.4 Certificate formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
10.5 Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
10.6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
10.7 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
10.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
11 Correlation clustering 160
11.1 Clustering formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
11.1.1 Linear programming relaxation . . . . . . . . . . . . . . . . . . . 162
11.1.2 Semidefinite programming relaxation . . . . . . . . . . . . . . . . 163
11.2 Learning formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
11.3 Dual formulation and kernels . . . . . . . . . . . . . . . . . . . . . . . . . 167
11.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.4.1 Irrelevant features . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.4.2 Email clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.4.3 Image segmentation . . . . . . . . . . . . . . . . . . . . . . . . . 170
11.5 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
11.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
xvi
IV Conclusions and future directions 175
12 Conclusions and future directions 176
12.1 Summary of contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 176
12.1.1 Structured maximum margin estimation . . . . . . . . . . . . . . . 177
12.1.2 Markov networks: max-margin, associative, relational . . . . . . . 180
12.1.3 Broader applications: parsing, matching, clustering . . . . . . . . . 182
12.2 Extensions and open problems . . . . . . . . . . . . . . . . . . . . . . . . 183
12.2.1 Theoretical analysis and optimization algorithms . . . . . . . . . . 183
12.2.2 Novel prediction tasks . . . . . . . . . . . . . . . . . . . . . . . . 184
12.2.3 More general learning settings . . . . . . . . . . . . . . . . . . . . 185
12.3 Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
A Proofs and derivations 188
A.1 Proof of Theorem 5.5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
A.1.1 Binary classification . . . . . . . . . . . . . . . . . . . . . . . . . 189
A.1.2 Structured classification . . . . . . . . . . . . . . . . . . . . . . . 190
A.2 AMN proofs and derivations . . . . . . . . . . . . . . . . . . . . . . . . . 195
A.2.1 Binary AMNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
A.2.2 Multi-class AMNs . . . . . . . . . . . . . . . . . . . . . . . . . . 197
A.2.3 Derivation of the factored primal and dual max-margin QP . . . . . 199
Bibliography 202
xvii
List of Tables
9.1 Parsing results on development set . . . . . . . . . . . . . . . . . . . . . . 140
9.2 Parsing results on test set . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
10.1 Bond connectivity prediction results . . . . . . . . . . . . . . . . . . . . . 156
xviii
List of Figures
1.1 Supervised learning setting . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Examples of complex prediction problems . . . . . . . . . . . . . . . . . . 4
2.1 Handwritten character recognition . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Classification loss and upper bounds . . . . . . . . . . . . . . . . . . . . . 20
3.1 Handwritten word recognition . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Chain Markov network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Diamond Markov network . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4 Diamond network junction tree . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 Marginal agreement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.6 Example parse tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1 Exact and approximate constraints for max-margin estimation . . . . . . . 51
4.2 A constraint generation algorithm. . . . . . . . . . . . . . . . . . . . . . . 52
5.1 Chain M
3
N example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2 Diamond Markov network . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.1 Block-coordinate ascent . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 SMO subproblem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.3 SMO pair selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.4 Structured SMO diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.5 Structured SMO pair selection . . . . . . . . . . . . . . . . . . . . . . . . 85
6.6 OCR results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
xix
7.1 Min-cut graph construction . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.2 α-expansion algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.3 Segbot: roving robot equipped with SICK2 laser sensors. . . . . . . . . . . 100
7.4 3D laser scan range map of the Stanford Quad. . . . . . . . . . . . . . . . 101
7.5 Comparison of terrain classification models (detail) . . . . . . . . . . . . . 102
7.6 Min-cut inference running times . . . . . . . . . . . . . . . . . . . . . . . 103
7.7 Comparison of terrain classification models (zoomed-out) . . . . . . . . . . 106
7.8 Labeled portion of the test terrain dataset (Ground truth and SVMpredictions)107
7.9 Labeled portion of the test terrain dataset (Voted-SVM and AMN predictions)108
8.1 Link model for document classification . . . . . . . . . . . . . . . . . . . 113
8.2 WebKB results: flat models . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8.3 WebKB results: link model . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.4 WebKB results: cocite model . . . . . . . . . . . . . . . . . . . . . . . . . 122
9.1 Two representations of a binary parse tree . . . . . . . . . . . . . . . . . . 130
10.1 PDB protein 1ANS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
10.2 Number of constraints vs. number of bonds . . . . . . . . . . . . . . . . . 150
10.3 Learning curve for bond connectivity prediction . . . . . . . . . . . . . . . 157
11.1 Learning to cluster with irrelevant features . . . . . . . . . . . . . . . . . . 169
11.2 Learning to organize email . . . . . . . . . . . . . . . . . . . . . . . . . . 170
11.3 Two segmentations by different users . . . . . . . . . . . . . . . . . . . . . 171
11.4 Two segmentations by different models . . . . . . . . . . . . . . . . . . . 172
xx
Chapter 1
Introduction
The breadth of tasks addressed by machine learning is rapidly expanding. Major appli-
cations include medical diagnosis, scientific discovery, financial analysis, fraud detection,
DNA sequence analysis, speech and handwriting recognition, game playing, image analy-
sis, robot locomotion and many more. Of course, the list of things we would like a computer
to learn to do is much, much longer. As we work our way down that list, we encounter the
need for very sophisticated decision making from our programs.
Some tasks, for example, handwriting recognition, are performed almost effortlessly
by a person, but remain difficult and error-prone for computers. The complex synthesis
of many levels of signal processing a person executes when confronted by a line of hand-
written text is daunting. The reconstruction of an entire sentence from the photons hitting
the retina off of each tiny patch of an image undoubtedly requires an elaborate interplay of
recognition and representation of the pen-strokes, the individual letters, whole words and
constituent phrases.
Computer scientists, as opposed to, say, neuroscientists, are primarily concerned with
achieving acceptable speed and accuracy of recognition rather than modeling this compli-
cated process with any biological verity. Computational models for handwriting recogni-
tion aim to capture the salient properties of the problem: typical shapes of the letters, likely
letter combinations that make up words, common ways to combine words into phrases, fre-
quent grammatical constructions of the phrases, etc. Machine learning offers an alternative
to encoding all the intricate details of such a model from scratch. One of its primary goals
1
2 CHAPTER 1. INTRODUCTION
is to devise efficient algorithms for training computers to automatically acquire effective
and accurate models from experience.
In this thesis, we present a discriminative learning framework and a novel family of effi-
cient models and algorithms for complex recognition tasks in several disciplines, including
natural language processing, computer vision and computational biology. We develop the-
oretical foundations for our approach and show a wide range of experimental applications,
including handwriting recognition, 3-dimensional terrain classification, disulfide connec-
tivity in protein structure prediction, hypertext categorization, natural language parsing,
email organization and image segmentation.
1.1 Supervised learning
The most basic supervised learning task is classification. Suppose we wish to learn to
recognize a handwritten character from a scanned image. This is a classification task,
because we must assigns a class (an English letter from ‘a’ through ‘z’) to an observation of
an object (an image). Essentially, a classifier is a function that maps an input (an image) to
an output (a letter). In the supervised learning setting, we construct a classifier by observing
labeled training examples, in our case, sample images paired with appropriate letters. The
main problem addressed by supervised learning is generalization. The learning program is
allowed to observe only a small sample of labeled images to produce an accurate classifier
on unseen images of letters.
More formally, let x denote an input. For example, a black-and-white image x can be
represented as a vector of pixel intensities. We use A to denote the space of all possible
inputs. Let y denote the output, and ¸ be the discrete space of possible outcomes (e.g.,
26 letters ‘a’-‘z’). A classifier (or hypothesis) h is a function from A to ¸, h : A → ¸.
We denote the set of all classifiers that our learning program can produce as H (hypothesis
class). Then given a set of labeled examples ¦x
(i)
, y
(i)
¦, i = 1, . . . , m, a learning program
seeks to produce a classifier h ∈ H that will work well on unseen examples x, usually by
finding h that accurately classifies training data. The diagram in Fig. 1.1 summarizes the
supervised learning setting.
1.1. SUPERVISED LEARNING 3
Labeled data
Learning
Prediction
Hypotheses
New data
Figure 1.1: Supervised learning setting
The problem of classification has a long history and highly developed theory and prac-
tice (see for example, Mitchell [1997]; Vapnik [1995]; Duda et al. [2000]; Hastie et al.
[2001]). The two most important dimensions of variation of classification algorithms is the
hypothesis class Hand the criterion for selection of a hypothesis h from Hgiven the train-
ing data. In this thesis, we build upon the generalized linear model family, which underlies
standard classifiers such as logistic regression and support vector machines. Through the
use of kernels to implicitly define high-dimensional and even infinite-dimensional input
representations, generalized linear models can approximate arbitrarily complex decision
boundaries.
The task of selecting a hypothesis h reduces to estimating model parameters. Broadly
speaking, probabilistic estimation methods associate a joint distribution p(x, y) or condi-
tional distribution p(y [ x) with h and select a model based on the likelihood of the data
[Hastie et al., 2001]. Joint distribution models are often called generative, while condi-
tional models are called discriminative. Large margin methods, by contrast, select a model
based on a more direct measure of confidence of its predictions on the training data called
the margin [Vapnik, 1995]. The difference between these two methods is one of the key
4 CHAPTER 1. INTRODUCTION
Thescreenwas
aseaofred
RSCCPCYWGGCPW
GQNCYPEGCSGPKV
brace
(a) (b) (c) (d)
Figure 1.2: Examples of complex prediction problems (inputs-top, outputs-bottom):
(a) handwriting recognition [image →word];
(b) natural language parsing [sentence →parse tree];
(c) disulfide bond prediction in proteins [amino-acid sequence →bond structure (shown in yellow)];
(d) terrain segmentation [3D image →segmented objects (trees, bushes, buildings, ground)]
themes in this thesis.
Most of the research has focused on the analysis and classification algorithms for the
case of binary outcomes [¸[ = 2, or a small number of classes. In this work, we focus
on prediction tasks that involve not a single decision with a small set of outcomes, but a
complex, interrelated collection of decisions.
1.2 Complex prediction problems
Consider once more the problem of character recognition. In fact, a more natural and useful
task is recognizing words and entire sentences. Fig. 1.2(a) shows an example handwritten
word “brace.” Distinguishing between the second letter and fourth letter (‘r’ and ‘c’) in iso-
lation is actually far from trivial, but in the context of the surrounding letters that together
form a word, this task is much less error-prone for humans and should be for computers
as well. It is also more complicated, as different decisions must be weighed against each
other to arrive at the globally satisfactory prediction. The space of all possible outcomes
1.2. COMPLEX PREDICTION PROBLEMS 5
¸ is immense, usually exponential in the number of individual decisions, for example, the
number of 5 letter sequences (26
5
). However, most of these outcomes are unlikely given
the observed input. By capturing the most salient structure of the problem, for example the
strong local correlations between consecutive letters, we will construct compact models
that efficiently deal with this complexity. Below we list several examples from different
fields.
• Natural language processing
Vast amounts of electronically available text have spurred a tremendous amount of
research into automatic analysis and processing of natural language. We mention
some of the lower-level tasks that have received a lot of recent attention [Charniak,
1993; Manning & Sch¨ utze, 1999]. Part-of-speech tagging involves assigning each
word in a sentence a part-of-speech tag, such as noun, verb, pronoun, etc. As with
handwriting recognition, capturing sequential structure of correlations between con-
secutive tags is key. In parsing, the goal is to recognize the recursive phrase structure
of a sentence, such as verbal, noun and prepositional phrases and their nesting in
relation to each other. Fig. 1.2(b) shows a parse tree corresponding to the sentence:
“The screen was a sea of red” (more on this in Ch. 9). Many other problems, such as
named-entity and relation extraction, text summarization, translation, involve com-
plex global decision making.
• Computational biology
The last two decades have yielded a wealth of high-throughput experimental data,
including complete sequencing of many genomes, precise measurements of protein
3D structure, genome-wide assays of mRNA levels and protein-protein interactions.
Major research has been devoted to gene-finding, alignment of sequences, protein
structure prediction, molecular pathway discovery [Gusfield, 1997; Durbin et al.,
1998]. Fig. 1.2(c) shows disulfide bond structure (shown in yellow) we would like to
predict from the amino-acid sequence of the protein (more on this in Ch. 10).
• Computer vision
As digital cameras and optical scanners become commonplace accessories, medical
imaging technology produces detailed physiological measurements, laser scanners
6 CHAPTER 1. INTRODUCTION
capture 3D environments, satellites and telescopes bring pictures of Earth and distant
stars, we are flooded with images we would like our computer to analyze. Example
tasks include object detection and segmentation, motion tracking, 3D reconstruction
from stereo and video, and much more [Forsyth & Ponce, 2002]. Fig. 1.2(d) shows a
3D laser range data image of the Stanford campus collected by a roving robot which
we would like to segment into objects such as trees, bushes, buildings, ground, etc.
(more on this in Ch. 7).
1.3 Structured models
This wide range of problems have been tackled using various models and methods. We
focus on the models that compactly capture correlation and constraint structure inherent to
many tasks. Abstractly, a model assigns a score (or likelihood in probabilistic models) to
each possible input/output pair (x, y), typically through a compact, parameterized scoring
function. Inference in these models refers to computing the highest scoring output given
the input and usually involves dynamic programming or combinatorial optimization.
• Markov networks
Markov networks (a.k.a. Markov random fields) are extensively used to model com-
plex sequential, spatial, and relational interactions in prediction problems arising in
many fields. These problems involve labeling a set of related objects that exhibit
local consistency. Markov networks compactly represent complex joint distributions
of the label variables by modeling their local interactions. Such models are encoded
by a graph, whose nodes represent the different object labels, and whose edges rep-
resent and quantify direct dependencies between them. The graphical structure of
the models encodes the qualitative aspects of the distribution: direct dependencies as
well as conditional independencies. The quantitative aspect of the model is defined
by the potentials that are associated with nodes and cliques of the graph. The graph-
ical structure of the network (more precisely, the treewidth of the graph, which we
formally define in Ch. 3) is critical to efficient inference and learning in the model.
• Context free grammars
1.3. STRUCTURED MODELS 7
Context-free grammars are one of the primary formalisms for capturing the recur-
sive structure of syntactic constructions [Manning & Sch¨ utze, 1999]. For example,
in Fig. 1.2, the non-terminal symbols (labels of internal nodes) correspond to syntac-
tic categories such as noun phrase (NP), verbal phrase (VP) or prepositional phrase
(PP) and part-of-speech tags like nouns (NN), verbs (VBD), determiners (DT) and
prepositions (IN). The terminal symbols (leaves) are the words of the sentence. A
CFG consists of recursive productions (e.g. V P → V P PP, DT → The) that
can be applied to derive a sentence of the language. The productions define the set
of syntactically allowed phrase structures (derivations). By compactly defining a
probability distribution over individual productions, probabilistic CFGs construct a
distribution over parse trees and sentences, and the prediction task reduces to finding
the most likely tree given the sentence. The context free restriction allows efficient
inference and learning in such models.
• Combinatorial structures
Many important computational tasks are formulated as combinatorial optimization
problems such as the maximum weight bipartite and perfect matching, spanning
tree, graph-cut, edge-cover, and many others [Lawler, 1976; Papadimitriou & Stei-
glitz, 1982; Cormen et al., 2001]. Although the term ‘model’ is often reserved for
probabilistic models, we use the term model very broadly, to include any scheme
that assigns scores to the output space ¸ and has a procedure for finding the opti-
mal scoring y. For example, the disulfide connectivity prediction in Fig. 1.2(c) can
be modeled by maximum weight perfect matchings, where the weights define po-
tential bond strength based on the local amino-acid sequence properties. The other
combinatorial structures we consider and apply in this thesis include graph cuts and
partitions, bipartite matchings, and spanning trees.
The standard methods of estimation for Markov networks and context free grammars
are based on maximum likelihood, both joint and conditional. However, maximum like-
lihood estimation of scoring function parameters for combinatorial structures is often in-
tractable because of the problem of defining a normalized distribution over an exponential
set of combinatorial structures.
8 CHAPTER 1. INTRODUCTION
1.4 Contributions
This thesis addresses the problemof efficient learning of high-accuracy models for complex
prediction problems. We consider a very large class of structured models, from Markov
networks to context free grammars to combinatorial graph structures such as matchings
and cuts. We focus on those models where exact inference is tractable, or can be efficiently
approximated.
◦ Learning framework for structured models
We propose a general framework for efficient estimation of models for structured
prediction. An alternative to likelihood-based methods, this framework builds upon
the large margin estimation principle. Intuitively, we find parameters such that in-
ference in the model (dynamic programming, combinatorial optimization) predicts
the correct answers on the training data with maximum confidence. We develop gen-
eral conditions under which exact large margin estimation is tractable and present
two formulations for structured max-margin estimation that define compact convex
optimization problems, taking advantage of prediction task structure. The first for-
mulation relies on the ability to express inference in the model as a compact convex
optimization problem. The second one only requires compactly expressing optimal-
ity of a given assignment according to the model and applies to a broader range of
combinatorial problems. These two formulations form the foundation which the rest
of the thesis develops.
◦ Markov networks
The largest portion of the thesis is devoted to novel estimation algorithms, represen-
tational extensions, generalization analysis and experimental validation for Markov
networks, a model class of choice in many structured prediction tasks in language,
vision and biology.
Low-treewidth Markov networks
We use graph decomposition to derive an exact, compact, convex learning for-
mulation for Markov networks with sequence and other low-treewidth structure.
The seamless integration of kernels with graphical models allows us to create
1.4. CONTRIBUTIONS 9
very rich models that leverage the immense amount of research in kernel de-
sign and graphical model decompositions for efficient, accurate prediction in
real-world tasks. We also use approximate graph decomposition to derive a
compact approximate formulation for Markov networks in which inference is
intractable.
Scalable online algorithm
We present an efficient algorithm for solving the estimation problem called
Structured SMO. Our online-style algorithm uses inference in the model and
analytic updates to solve extremely large estimation problems.
Generalization analysis
We analyze the theoretical generalization properties of max-margin estimation
in Markov networks and derive a novel margin-based bound for structured pre-
diction. This bound is the first to address structured error (e.g. proportion
of mislabeled pixels in an image) and uses a proof that exploits the graphical
model structure.
Learning associative Markov networks (AMNs)
We define an important subclass of Markov networks that captures positive cor-
relations present in many domains. We show that for AMNs over binary vari-
ables, our framework allows exact estimation of networks of arbitrary connec-
tivity and topology, for which likelihood methods are believed to be intractable.
For the non-binary case, we provide an approximation that works well in prac-
tice. We present an AMN-based method for object segmentation from 3D range
data. By constraining the class of Markov networks to AMNs, our models are
learned efficiently and, at run-time, scale up to tens of millions of nodes and
edges.
Representation and learning of relational Markov networks
We introduce relational Markov networks (RMNs), which compactly define
templates for Markov networks for domains with relational structure objects,
10 CHAPTER 1. INTRODUCTION
attributes, relations. The graphical structure of an RMN is based on the rela-
tional structure of the domain, and can easily model complex interaction pat-
terns over related entities. We use approximate inference in these complex mod-
els, in which exact inference is intractable, to derive an approximate learning
formulation. We apply this class of models to classification of hypertext using
hyperlink structure to define relations between webpages.
◦ Broader applications: parsing, matching, clustering
The other large portion the thesis addresses a range of prediction tasks with very di-
verse models: context free grammars for natural language parsing, perfect matchings
for disulfide connectivity in protein structure prediction, graph partitions for cluster-
ing documents and segmenting images.
Learning to parse
We exploit context free grammar structure to derive a compact max-margin
formulation and show high-accuracy parsing in cubic time by exploiting novel
kinds of lexical information. We show experimental evidence of the model’s
improved performance over several baseline models.
Learning to match
We use combinatorial properties of weighted matchings to develop an exact,
efficient algorithm for learning to match. We apply our framework to predic-
tion of disulfide connectivity in proteins using perfect non-bipartite matchings.
The algorithm we propose uses kernels, which makes it possible to efficiently
embed the features in very high-dimensional spaces and achieve state-of-the-art
accuracy.
Learning to cluster
We derive a max-margin formulation for learning the affinity metric for clus-
tering from clustered training data. In contrast to algorithms that learn a metric
independently of the algorithm that will be used to cluster the data, we describe
a formulation that tightly integrates metric learning with the clustering algo-
rithm, tuning one to the other in a joint optimization. Experiments on synthetic
and real-world data show the ability of the algorithm to learn an appropriate
1.5. THESIS OUTLINE 11
clustering metric for a variety of desired clusterings, including email folder or-
ganization and image segmentation.
1.5 Thesis outline
Below is a summary of the rest of the chapters in the thesis:
Chapter 2. Supervised learning: We review basic definitions and statistical framework
for classification. We define hypothesis classes, loss functions, risk. We consider
generalized linear models, including logistic regression and support vector machines,
and review estimation methods based on maximizing likelihood, conditional likeli-
hood and margin. We describe the relationship between the dual estimation problems
and kernels.
Chapter 3. Structured models: In this chapter, we define the abstract class of structured
prediction problems and models addressed by the thesis. We compare probabilistic
models, generative and discriminative and unnormalized models. We describe repre-
sentation and inference for Markov networks, including dynamic and linear program-
ming inference. We also briefly describe context free grammars and combinatorial
structures as models.
Chapter 4. Structured maximum margin estimation: This chapter outlines the main prin-
ciples of maximum margin estimation for structured models. We address the expo-
nential blow-up of the naive problem formulation by deriving two general equivalent
convex formulation. These formulations, min-max and certificate, allow us to ex-
ploit decomposition and combinatorial structure of the prediction task. They lead
to polynomial size programs for estimation of models where the prediction problem
is tractable. We also discuss approximations, in particular using upper and lower
bounds, for solving intractable or very large problems.
Chapter 5. Markov networks: We review maximum conditional likelihood estimation
and present maximum margin estimation for Markov networks. We use graphical
model decomposition to derive a convex, compact formulation that seamlessly in-
tegrates kernels with graphical models. We analyze the theoretical generalization
12 CHAPTER 1. INTRODUCTION
properties of max-margin estimation and derive a novel margin-based bound for
structured classification.
Chapter 6. M
3
N algorithms and experiments: We present an efficient algorithmfor solv-
ing the estimation problem in graphical models, called Structured SMO. Our online-
style algorithm uses inference in the model and analytic updates to solve extremely
large quadratic problems. We present experiments with handwriting recognition,
where our models significantly outperform other approaches by effectively capturing
correlation between adjacent letters and incorporating high-dimensional input repre-
sentation via kernels.
Chapter 7. Associative Markov networks: We define an important subclass of Markov
networks, associative Markov networks (AMNs), that captures positive interactions
present in many domains. We show that for associative Markov networks of over bi-
nary variables, max-margin estimation allows exact training of networks of arbitrary
connectivity and topology, for which maximum likelihood methods are believed to
be intractable. For the non-binary case, we provide an approximation that works
well in practice. We present an AMN-based method for object segmentation from
3D range data that scales to very large prediction tasks involving tens of millions of
points.
Chapter 8. Relational Markov networks: We introduce the framework of relational Mar-
kov networks (RMNs), which compactly defines templates for Markov networks in
domains with rich structure modeled by objects, attributes and relations. The graph-
ical structure of an RMN is based on the relational structure of the domain, and can
easily model complex patterns over related entities. As we show, the use of an undi-
rected, discriminative graphical model avoids the difficulties of defining a coherent
generative model for graph structures in directed models and allows us tremendous
flexibility in representing complex patterns. We provide experimental results on a
webpage classification task, showing that accuracy can be significantly improved by
modeling relational dependencies.
Chapter 9. Context free grammars: We present max-margin estimation for natural lan-
guage parsing on the decomposition properties of context free grammars. We show
1.6. PREVIOUSLY PUBLISHED WORK 13
that this framework allows high-accuracy parsing in cubic time by exploiting novel
kinds of lexical information. We show experimental evidence of the model’s im-
proved performance over several baseline models.
Chapter 10. Perfect matchings: We apply our framework to learning to predict disulfide
connectivity in proteins using perfect matchings. We use combinatorial properties of
weighted matchings to develop an exact, efficient algorithm for learning the param-
eters of the model. The algorithm we propose uses kernels, which makes it possible
to efficiently embed the features in very high-dimensional spaces and achieve state-
of-the-art accuracy.
Chapter 11. Correlation clustering: In this chapter, we derive a max-margin formula-
tion for learning affinity scores for correlation clustering from clustered training data.
We formulate the approximate learning problem as a compact convex program with
quadratic objective and linear or positive-semidefinite constraints. Experiments on
synthetic and real-world data show the ability of the algorithm to learn an appro-
priate clustering metric for a variety of desired clusterings, including email folder
organization and image segmentation.
Chapter 12. Conclusions and future directions: We reviewthe main contributions of the
thesis and summarize their significance, applicability and limitations. We discuss ex-
tensions and future research directions not addressed in the thesis.
1.6 Previously published work
Some of the work described in this thesis has been published in conference proceedings.
The min-max and certificate formulations for structured max-margin estimation have not
been published in their general form outlined in Ch. 4, although they underly several pa-
pers mentioned below. The polynomial formulation of maximum margin Markov networks
presented in Ch. 5 was published for a less general case, using a dual decomposition tech-
nique [Taskar et al., 2003a]. Work on associative Markov networks (Ch. 7) was published
with experiments on hypertext and news-wire classification [Taskar et al., 2004a]. A paper
14 CHAPTER 1. INTRODUCTION
on 3D object segmentation using AMNs, which presents a experiments on terrain classifi-
cation and other tasks, is currently under review (joint work with Drago Anguelov, Vassil
Chatalbashev, Dinkar Gupta, Geremy Heitz, Daphne Koller and Andrew Ng). Taskar et al.
[2002] and Taskar et al. [2003b] defined and applied the Relational Markov networks
(Ch. 8), using maximum (conditional) likelihood estimation. Natural language parsing
in Ch. 9 was published in Taskar et al. [2004b]. Disulfide connectivity prediction using
perfect matchings in Ch. 10 (joint work with Vassil Chatalbashev and Daphne Koller) is
currently under review. Finally, work on correlation clustering in Ch. 11, done jointly with
Pieter Abbeel and Andrew Ng, has not been published.
Part I
Models and methods
15
Chapter 2
Supervised learning
In supervised learning, we seek a function h : A → ¸ that maps inputs x ∈ A to outputs
y ∈ ¸. The input space A is an arbitrary set (often A = IR
n
), while the output space ¸
we consider in this chapter discrete. A supervised learning problem with discrete outputs,
¸ = ¦y
1
, . . . , y
k
¦, where k is the number of classes, is called classification. In handwritten
character recognition, for example, A is the set of images of letters and ¸ is the alphabet
(see Fig. 2.1).
The input to an algorithm is training data, a set of m i.i.d. (independent and identically
distributed) samples S = ¦(x
(i)
, y
(i)

m
i=1
drawn from a fixed but unknown distribution D
over A ¸. The goal of a learning algorithm is to output a hypothesis h such that h(x)
will approximate y on new samples from the distribution (x, y) ∼ D.
Learning algorithms can be distinguished among several dimensions, chief among them
is the hypothesis class H of functions h the algorithm outputs. Numerous classes of func-
tions have been well studied, including decision trees, neural networks, nearest-neighbors,
generalized log-linear models and kernel methods (see Quinlan [2001]; Bishop [1995];
Hastie et al. [2001]; Duda et al. [2000], for in-depth discussion of these and many other
models). We will concentrate on the last two classes, for several reasons we discuss be-
low, including accuracy, efficiency, and extensibility to more complex structured prediction
tasks will consider in the next chapter.
The second crucial dimension of a learning algorithm is the criterion for selection of h
fromH. We arrive at such a criterion by quantifying what it means for h(x) to approximate
16
17
y. The risk functional 1

D
[(h)] measures the expected error of the approximation:
1

D
[h] = E
(x,y)∼D
[(x, y, h(x))], (2.1)
where the loss function : A ¸ ¸ → IR
+
measures the penalty for predicting h(x)
on the sample (x, y). In general, we assume that (x, y, ˆ y) = 0 if y = ˆ y.
A common loss function for classification is 0/1 loss

0/1
(x, y, h(x)) ≡ 1I(y ,= h(x)),
where 1I() denotes the indicator function, that is, 1I(true) = 1 and 1I(false) = 0.
Since we do not generally know the distribution D, we estimate the risk of h using its
empirical risk 1

S
, computed on the training sample S:
1

S
[h] =
1
m
m

i=1
(x
(i)
, y
(i)
, h(x
(i)
)) =
1
m
m

i=1

i
(h(x
(i)
)), (2.2)
where we abbreviate (x
(i)
, y
(i)
, h(x
(i)
)) =
i
(h(x
(i)
)). For 0/1 loss, 1

S
[h] is simply the
proportion of training examples that h misclassifies. 1

S
[h] is often called the training
error or training loss.
If our set of hypotheses, H, is large enough, we will be able to find h that has zero or
very small empirical risk. However, simply selecting a hypothesis with lowest risk
h

= arg min
h∈1
1

S
[h],
is generally not a good idea. For example, if A = IR, ¸ = IR and Hincludes all polynomi-
als of degree m−1, we can always find a polynomial h that passes through all the sample
points (x
(i)
, y
(i)
), i = (1, ..., m) assuming that all the x
(i)
are unique. This polynomial is
very likely to overfit the training data, that is, it will have zero empirical risk, but high ac-
tual risk. The key to selecting a good hypothesis is to trade-off complexity of class H (e.g.
the degree of the polynomial) with the error on the training data as measured by empirical
risk 1

S
. For a vast majority of supervised learning algorithms, this fundamental balance is
18 CHAPTER 2. SUPERVISED LEARNING
achieved by minimizing the weighted combination of the two criteria:
h

= arg min
h∈1
_
T[h] + C1

S
[h]
_
, (2.3)
where T[h] measures the inherent dimension or complexity of h, and C ≥ 0 is a trade-
off parameter. We will not go into derivation of various complexity measures T[h] here,
but simply adopt the standard measures as needed and refer the reader to Vapnik [1995];
Devroye et al. [1996]; Hastie et al. [2001] for details. The term T[h] is often called
regularization.
Depending on the complexity of the class H, the search for the optimal h

in (2.3)
may be a daunting task
1
. For many classes, for example decision trees and multi-layer
neural networks, it is intractable [Bishop, 1995; Quinlan, 2001], and we must resort to
approximate, greedy optimization methods. For these intractable classes, the search pro-
cedure used by the learning algorithm is crucial. Below however, we will concentrate on
models where the optimal h

can be found efficiently using convex optimization in poly-
nomial time. Hence, the learning algorithms we consider are completely characterized by
the hypothesis class H, the loss function , and the regularization T[h].
In general, we consider hypothesis classes of the following parametric form:
h
w
(x) = arg max
y∈¸
f(w, x, y), (2.4)
where f(w, x, y) is a function f : ¼A ¸ → IR, where w ∈ ¼ is a set of parameters,
usually with ¼ ⊆ IR
n
. We assume that ties in the arg max are broken using some arbitrary
but fixed rule. As we discuss below, this class of hypotheses is very rich and includes
many standard models. The formulation in (2.4) of the hypothesis class in terms of an
optimization procedure will become crucial to extending supervised learning techniques to
cases where the output space ¸ is more complex.
1
For classification, minimizing the objective with the usual 0/1 training error is generally a very difficult
problem with multiple maxima for most realistic H. See discussion in the next section about approaches to
dealing with 0/1 loss.
2.1. CLASSIFICATION WITH GENERALIZED LINEAR MODELS 19
a b c d e
Figure 2.1: Handwritten character recognition: sample letters from Kassel [1995] data set.
2.1 Classification with generalized linear models
For classification, we consider the generalized linear family of hypotheses H. Given n
real-valued basis functions f
j
: A ¸ → IR, a hypothesis h
w
∈ H is defined by a set of n
coefficients w
j
∈ IR such that:
h
w
(x) = arg max
y∈¸
n

i=1
w
j
f
j
(x, y) = arg max
y∈¸
w
¯
f (x, y). (2.5)
Consider the character recognition example in Fig. 2.1. Our input x is a vector of
pixel values of the image and y is the alphabet ¦a, . . . , z¦. We might have a basis function
f
j
(x, y) = 1I(x
row,col
= on ∧ y = char) for each possible (row, col) and char ∈ ¸,
where x
row,col
denotes the value of pixel (row, col). Since different letters tend to have
different pixels turned on, this very simple model captures enough information to perform
reasonably well.
The most common loss for classification is 0/1 loss. Minimizing the 0/1 risk is generally
a very difficult problem with multiple maxima for any large class H. The standard solution
is minimizing an upper bound on the 0/1 loss, (x, y, h(x)) ≥ (x, y, h(x)). (In addition
to computational advantages of this approach, there are statistical benefits of minimizing a
convex upper bound [Bartlett et al., 2003]). Two of the primary classification methods we
consider, logistic regression and support vector machines, differ primarily in their choice of
the upper bound on the training 0/1 loss. The regularization T[h
w
] for the linear family is
typically the norm of the parameters [[w[[
p
for p = 1, 2. Intuitively, a zero, or small weight
w
j
implies that the hypothesis h
w
does not depend on the value of f
j
(x, y) and hence is
simpler than a h
w
with a large weight w
j
.
20 CHAPTER 2. SUPERVISED LEARNING
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
0
0.5
1
1.5
2
2.5
3
3.5
log−loss
hinge−loss
0/1−loss
Figure 2.2: 0/1-loss upper bounded by log-loss and hinge-loss. Horizontal axis shows
w
¯
f (x, y) −max
y

,=y
w
¯
f (x, y
t
), where y is the correct label for x, while the vertical axis
show the value of the associated loss. The log-loss is shown up to an additive constant for
illustration purposes.
2.2 Logistic regression
In logistic regression, we assign a probabilistic interpretation to the hypothesis h
w
as defin-
ing a conditional distribution:
P
w
(y [ x) =
1
Z
w
(x)
exp¦w
¯
f (x, y)¦, (2.6)
where Z
w
(x) =

y∈¸
exp¦w
¯
f (x, y)¦. The optimal weights are selected by maximiz-
ing the conditional likelihood of the data (minimizing the log-loss) with some regulariza-
tion. This approach is called the (regularized) maximum likelihood estimation. Common
choices for regularization are 1 or 2-norm regularization on the weights; we use 2-norm
below:
min
1
2
[[w[[
2
+ C

i
log Z
w
(x
(i)
) −w
¯
f (x
(i)
, y
(i)
), (2.7)
where C is a user-specified constant the determines the trade-off between regularization
and likelihood of the data. The log-loss log Z
w
(x) −w
¯
f (x, y) is an upper bound (up to a
constant) on the 0/1 loss
0/1
(see Fig. 2.2).
2.3. LOGISTIC DUAL AND MAXIMUM ENTROPY 21
2.3 Logistic dual and maximum entropy
The objective function is convex in the parameters w, so we have an unconstrained (differ-
entiable) convex optimization problem. The gradient with respect to w is given by:
w+ C

i
E
i,w
[f
i
(x
(i)
, y)] −f
i
(x
(i)
, y
(i)
) = w−C

i
E
i,w
[∆f
i
(y)],
where E
i,w
[f(y)] =

y
f(y)P
w
(y [ x
(i)
) is the expectation under the conditional distribu-
tion P
w
(y [ x
(i)
) and ∆f
i
(y) = f (x
(i)
, y
(i)
) − f (x
(i)
, y). Ignoring the regularization term,
the gradient is zero when the basis function expectations are equal to the basis functions
evaluated on the labels y
(i)
. It can be shown [Cover & Thomas, 1991] that the dual of the
maximum likelihood problem (without regularization) is the maximum entropy problem:
max −

i,y
P
w
(y [ x
(i)
) log P
w
(y [ x
(i)
) (2.8)
s.t. E
i,w
[∆f
i
(y)] = 0, ∀i.
We can interpret logistic regression as trying to match the empirical basis function expec-
tations while maintaining a high entropy conditional distribution P
w
(y [ x).
2.4 Support vector machines
Support vector machines [Vapnik, 1995] select the weights based on the “margin” of con-
fidence of h
w
. In the multi-class SVM formulation [Weston & Watkins, 1998; Crammer &
Singer, 2001], the margin on example i quantifies by how much the true label “wins” over
the wrong ones:
γ
i
=
1
[[w[[
min
y,=y
(i)
w
¯
f (x
(i)
, y
(i)
) −w
¯
f (x
(i)
, y) =
1
[[w[[
min
y,=y
(i)
w
¯
∆f
i
(y),
22 CHAPTER 2. SUPERVISED LEARNING
where ∆f
i
(y) = f (x
(i)
, y
(i)
) −f (x
(i)
, y). Maximizing the smallest such margin (and allow-
ing for negative margins) is equivalent to solving the following quadratic program:
min
1
2
[[w[[
2
+ C

i
ξ
i
(2.9)
s.t. w
¯
∆f
i
(y) ≥
0/1
(y) −ξ
i
, ∀i, ∀y ∈ ¸.
Note that the slack variable ξ
i
is constrained to be positive in the above program since
w
¯
∆f
i
(y
(i)
) = 0 and
0/1
(y
(i)
) = 0. We can also express ξ
i
as max
y

0/1
i
(y) −w
¯
∆f
i
(y),
and the optimization problem Eq. (2.9) in a form similar to Eq. (2.7):
min
1
2
[[w[[
2
+ C

i
max
y
[
0/1
i
(y) −w
¯
∆f
i
(y)]. (2.10)
The hinge-loss max
y
[
0/1
i
(y) − w
¯
∆f
i
(y)] is also an upper bound on the 0/1 loss
0/1
(see Fig. 2.2).
2.5 SVM dual and kernels
The form of the dual of Eq. (2.9) is crucial to efficient solution of SVM and the ability to
use a high or even infinite dimensional set of basis functions via kernels.
max

i,y
α
i
(y)
0/1
i
(y) −
1
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,y
α
i
(y)∆f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(2.11)
s.t.

y
α
i
(y) = C, ∀i; α
i
(y) ≥ 0, ∀i, y.
In the dual, the α
i
(y) variables correspond to the w
¯
∆f
i
(y) ≥
0/1
(y)−ξ
i
constraints in the
primal Eq. (2.9). The solution to the dual α

gives the solution to the primal as a weighted
combination of basis functions of examples:
w

=

i,y
α

i
(y)∆f
i
(y).
2.5. SVM DUAL AND KERNELS 23
The pairings of examples and incorrect labels, (i, y), that have non-zero α

i
(y), are called
support vectors.
An important feature of the dual formulation is that the basis functions f appear only as
dot products. Expanding the quadratic term, we have:
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,y
α
i
(y)∆f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
=

i,y

j,¯ y
α
i
(y)α
j
(¯ y)∆f
i
(y)
¯
∆f
j
(¯ y).
Hence, as long as the dot product f (x, y)
¯
f (¯ x, ¯ y) can be computed efficiently, we can
solve Eq. (2.11) independently of the actual dimension of f . Note that at classification
time, we also do not need to worry about the dimension of f since:
w
¯
f (x, ¯ y) =

i,y
α
i
(y)∆f
i
(y)
¯
f (x, ¯ y) =

i,y
α
i
(y)[f (x
(i)
, y
(i)
)
¯
f (x, ¯ y)−f (x
(i)
, y)
¯
f (x, ¯ y)].
For example, we might have basis functions that are polynomial of degree d in terms of
image pixels, f
j
(x, y) = 1I(x
row
1
,col
1
= on ∧ . . . ∧ x
row
d
,col
d
= on ∧ y = char) for each
possible (row
1
, col
1
) . . . (row
d
, col
d
) and char ∈ ¸. Computing this polynomial kernel
can be done independently of the dimension d, even though the number of basis functions
grows exponentially with d [Vapnik, 1995].
In fact, logistic regression can also be kernelized. However, the hinge loss formulation
usually produces sparse solutions in terms of the number of support vectors, while solutions
to the corresponding kernelized log-loss problem are generally non-sparse (all examples
are support vectors) and require approximations for even relatively small datasets [Wahba
et al., 1993; Zhu & Hastie, 2001].
Chapter 3
Structured models
Consider once more the problem of character recognition. In fact, a more natural and useful
task is recognizing words and entire sentences. Fig. 3.1 shows an example handwritten
word “brace.” Distinguishing between the second letter and fourth letter (‘r’ and ‘c’) in
isolation is far from trivial, but in the context of the surrounding letters that together form
a word, this task is much less error-prone for humans and should be for computers as well.
In this chapter, we consider prediction problems in which the output is not a single
discrete value y, but a set of values y = (y
1
, . . . , y
L
), for example an entire sequence
of L characters. For concreteness, let the number of variables L be fixed. The output
space ¸ ⊆ ¸
1
. . . ¸
L
we consider is a subset of product of output spaces of single
variables. In word recognition, each ¸
j
is the alphabet, while ¸ is the dictionary. This
joint output space is often a proper subset of the product of singleton output spaces, ¸ ⊂
¸
1
. . .¸
L
. In word recognition, we might restrict that the letter ‘q’ never follows by ‘z’ in
English. In addition to “hard” constraints, the output variables are often highly correlated,
e.g. consecutive letters in a word. We refer to joint spaces with constraints and correlations
as structured. We call problems with discrete output spaces structured classification or
structured prediction. Structured models we consider in this chapter (and thesis) predict
the outputs jointly, respecting the constraints and exploiting the correlations in the output
space.
The range of prediction problems these broad definitions encompass is immense, aris-
ing in fields as diverse as natural language analysis, machine vision, and computational
24
3.1. PROBABILISTIC MODELS: GENERATIVE AND CONDITIONAL 25
b r a c e
Figure 3.1: Handwritten word recognition: sample from Kassel [1995] data set.
biology, to name a few. The class of structured models H we consider is essentially of the
same form as in previous chapter, except that y has been replaced by y:
h
w
(x) = arg max
y : g(x,y)≤0
w
¯
f (x, y), (3.1)
where as before f (x, y) is a vector of functions f : A ¸ → IR
n
. The output space
¸ = ¦y : g(x, y) ≤ 0¦ is defined using a vector of functions g(x, y) that define the
constraints, where g : A ¸ → IR
k
. This formulation is very general. Clearly, for
many f , g pairs, finding the optimal y is intractable. For the most part, we will restrict our
attention to models where this optimization problemcan be solved in polynomial time. This
includes, for example, probabilistic models like Markov networks (in certain cases) and
context-free grammars, combinatorial optimization problems like min-cut and matching,
convex optimization such as linear, quadratic and semi-definite programming. In other
cases, like intractable Markov networks (Ch. 8) and correlation clustering (Ch. 11), we use
an approximate polynomial time optimization procedure.
3.1 Probabilistic models: generative and conditional
The term model is often reserved for probabilistic models, which can be subdivided into
generative and conditional with respect to the prediction task. A generative model assigns
a normalized joint density p(x, y) to the input and output space A ¸ with
p(x, y) ≥ 0,

y∈¸
_
x∈.
p(x, y) = 1.
26 CHAPTER 3. STRUCTURED MODELS
A conditional model assigns a normalized density p(y [ x) only over the output space ¸
with
p(y [ x) ≥ 0,

y∈¸
p(y [ x) = 1 ∀x ∈ A.
Probabilistic interpretation of the model offers well-understood semantics and an im-
mense toolbox of methods for inference and learning. It also provides an intuitive measure
of confidence in the predictions of a model in terms of conditional probabilities. In addi-
tion, generative models are typically structured to allow very efficient maximum likelihood
learning. A very common class of generative models is the exponential family:
p(x, y) ∝ exp¦w
¯
f (x, y)¦.
For exponential families, the maximum likelihood parameters w with respect to the joint
distribution can be computed in closed form using the empirical basis function expectations
E
S
[f (x, y)] [DeGroot, 1970; Hastie et al., 2001].
Of course, this efficiency comes at a price. Any model is an approximation to the true
distribution underlying the data. A generative model must make simplifying assumptions
(more precisely, independence assumptions) about the entire p(x, y), while a conditional
model makes many fewer assumption by focusing on p(y [ y). Because of this, by opti-
mizing the model to fit the joint distribution p(x, y), we may be tuning the approximation
away from optimal conditional distribution p(y [ x), which we use to make the predictions.
Given sufficient data, the conditional model will learn the best approximation to p(y [ x)
possible using w, while the generative model p(x, y) will not necessarily do so. Typically,
however, generative models actually need fewer samples to converge to a good estimate of
the joint distribution than conditional models need to accurately represent the conditional
distribution. In a regime with very few training samples (relative to the number of param-
eters w), generative models may actually outperform conditional models [Ng & Jordan,
2001].
3.2. PREDICTION MODELS: NORMALIZED AND UNNORMALIZED 27
3.2 Prediction models: normalized and unnormalized
Probabilistic semantics are certainly not necessary for a good predictive model if we are
simply interested in the optimal prediction (the arg max in Eq. (3.1)). As we discussed
in the previous chapter, support vector machines, which do not represent a conditional
distribution, typically perform as well or better than logistic regression [Vapnik, 1995;
Cristianini & Shawe-Taylor, 2000].
In general, we can often achieve higher accuracy models when we do not learn a nor-
malized distribution over the outputs, but concentrate on the margin or decision boundary,
the difference between the optimal y and the rest. Even more importantly, in many cases
we discuss below, normalizing the model (summing over the entire ¸) is intractable, while
the optimal y can be found in polynomial time. This fact makes standard maximum like-
lihood estimation infeasible. The learning methods we advocate in this thesis circumvent
this problem by requiring only the maximization problem to be tractable. We still heav-
ily rely on the representation and inference tools familiar from probabilistic models for
the construction of and prediction in unnormalized models, but largely dispense with the
probabilistic interpretation when needed. Essentially, we use the term model very broadly,
to include any scheme that assigns scores to the output space ¸ and has a procedure for
finding the optimal scoring y.
In this chapter, we review basic concepts in probabilistic graphical models called Mar-
kov networks or Markov randomfields. We also briefly touch upon examples of context-free
grammars and combinatorial problems that will be explained in greater detail in Part III to
illustrate the range of prediction problems we address.
3.3 Markov networks
Markov networks provide a framework for a rich family of models for both discrete and
continuous prediction [Pearl, 1988; Cowell et al., 1999]. The models treat the inputs and
outputs as random variables X with domain A and Y with domain ¸ and compactly de-
fine a conditional density p(Y [ X) or distribution P(Y [ X) (we concentrate here on the
28 CHAPTER 3. STRUCTURED MODELS
conditional Markov networks or CRFs [Lafferty et al., 2001]). The advantage of a graphi-
cal framework is that it can exploit sparseness in the correlations between outputs Y . The
graphical structure of the models encodes the qualitative aspects of the distribution: direct
dependencies as well as conditional independencies. The quantitative aspect of the model
is defined by the potentials that are associated with nodes and cliques of the graph. Before
a formal definition, consider a first-order Markov chain a model for the word recognition
task. In Fig. 3.2, the nodes are associated with output variables Y
i
and the edges correspond
to direct dependencies or correlations. We do not explicitly represent the inputs X in the
figure. For example, the model encodes that Y
j
is conditionally independent of the rest of
the variables given Y
j−1
, Y
j+1
. Intuitively, adjacent letters in a word are highly correlated,
but the first-order model is making the assertion (which is certainly an approximation) that
once the value of a letter Y
j
is known, the correlation between a letter Y
b
before j and a
letter Y
a
after j is negligible. More precisely, we use a model where
P(Y
b
[ Y
j
, Y
a
, x) = P(Y
b
[ Y
j
, x), P(Y
a
[ Y
j
, Y
b
, x) = P(Y
a
[ Y
j
, x), b < j < a.
For the purposes of finding the most likely y, this conditional independence property means
that the optimization problem is decomposable: given that Y
j
= y
j
, it suffices to separately
find the optimal subsequence from 1 to j ending with y
j
, and the optimal subsequence
starting with y
j
from j to L.
3.3.1 Representation
The structure of a Markov network is defined by an undirected graph ( = (1, c), where
the nodes are associated with variables 1 = ¦Y
1
, . . . , Y
L
¦. A clique is a set of nodes c ⊆ 1
that form a fully connected subgraph (every two nodes are connected by an edge). Note that
each subclique of a clique is also a clique, and we consider each node a singleton clique.
In the chain network in Fig. 3.2, the cliques are simply the nodes and the edges: ((() =
¦¦Y
1
¦, . . . , ¦Y
5
¦, ¦Y
1
, Y
2
¦, . . . , ¦Y
4
, Y
5
¦¦. We denote the set of variables in a clique c as
Y
c
, an assignment of variables in the clique as y
c
and the space of all assignments to
the clique as ¸
c
. We focus on discrete output spaces ¸ below, but many of the same
representation and inference concepts translate to continuous domains. No assumption is
3.3. MARKOV NETWORKS 29
Figure 3.2: First-order Markov chain: φ
i
(Y
i
) are node potentials, φ
i,i+1
(Y
i
, Y
i+1
) are edge
potentials (dependence on x is not shown).
made about A.
Definition 3.3.1 A Markov network is defined by an undirected graph ( = (1, c) and a
set of potentials Φ = ¦φ
c
¦. The nodes are associated with variables 1 = ¦Y
1
, . . . , Y
L
¦.
Each clique c ∈ ((() is associated with a potential φ
c
(x, y
c
) with φ
c
: A ¸
c
→ IR
+
,
which specifies a non-negative value for each assignment y
c
to variables in Y
c
and any
input x. The Markov network ((, Φ) defines a conditional distribution:
P(y [ x) =
1
Z(x)

c∈((Ç)
φ
c
(x, y
c
),
where ((() is the set of all the cliques of the graph and Z(x) is the partition function
given by Z(x) =

y∈¸

c∈((Ç)
φ
c
(x, y
c
).
In our example Fig. 3.2, we have node and edge potentials. Intuitively, the node poten-
tials quantify the correlation between the input x and the value of the node, while the edge
potentials quantify the correlation between the pair of adjacent output variables as well as
the input x. Potentials do not have a local probabilistic interpretation, but can be thought
of as defining an unnormalized score for each assignment in the clique. Conditioned on
the image input, appropriate node potentials in our network should give high scores to the
correct letters (‘b’,‘r’,‘a’,‘c’,‘e’), though perhaps there would be some ambiguity with the
second and fourth letter. For simplicity, assume that the edge potentials would not depend
30 CHAPTER 3. STRUCTURED MODELS
on the images, but simply should give high scores to pairs of letters that tend to appear often
consecutively. Multiplied together, these scores should favor the correct output “brace”.
In fact, a Markov network is a generalized log-linear model, since the potentials φ
c
(x
c
, y
c
)
could be represented (in log-space) as a sum of basis functions over x, y
c
:
φ
c
(x
c
, y
c
) = exp
_
n
c

k=1
w
c,k
f
c,k
(x, y
c
)
_
= exp
_
w
¯
c
f
c
(x, y
c
)
¸
where n
c
is the number of basis functions for the clique c. Hence the log of the conditional
probability is given by:
log P(y [ x) =

c∈((Ç)
w
¯
c
f
c
(x, y
c
) −log Z
w
(x).
In case of node potentials for word recognition, we could use the same basis functions as
for individual character recognition: f
j,k
(x, y
j
) = 1I(x
j,row,col
= on ∧ y
j
= char) for each
possible (row, col) in x
j
, the window of the image that corresponds to letter j and each
char ∈ ¸
j
(we assume the input has been segmented into images x
j
that correspond to
letters). In general, we condition a clique only on a portion of the input x, which we denote
as x
c
. For the edge potentials, we can define basis functions for each combination of letters
(assume for simplicity no dependence on x) : f
j,j+1,k
(x, y
j
, y
j+1
) = 1I(y
j
= char
1
∧y
j+1
=
char
2
) for each char
1
∈ ¸
j
and char
2
∈ ¸
j+1
. In this problem (as well as many others),
we are likely to “tie” or “share” the parameters of the model w
c
across cliques. Usually, all
single node potentials would share the same weights and basis functions (albeit the relevant
portion of the input x
c
is different) and similarly for the pairwise cliques, no matter in what
position they appear in the sequence.
1
With slight abuse of notation, we stack all basis functions into one vector f . For the
sequence model, f has node functions and edge functions, so when c is a node, the edge
functions in f (x
c
, y
c
) are defined to evaluate to zero. Similarly, when c is an edge, the node
1
Sometimes we might actually want some dependence on the position in the sequence, which can be
accomplished by adding more basis functions that condition on the position of the clique.
3.3. MARKOV NETWORKS 31
functions in f (x
c
, y
c
) are also defined to evaluate to zero. Now we can write:
f (x, y) =

c∈((Ç)
f (x
c
, y
c
).
We stack the weights in the corresponding manner, so the most likely assignment according
to the model is given by:
arg max
y∈¸
log P
w
(y [ x) = arg max
y∈¸
w
¯
f (x, y),
in the same form as Eq. (3.1).
3.3.2 Inference
There are several important questions that can be answered by probabilistic models. The
task of finding the most likely assignment, known as maximum a-posteriori (MAP) or most
likely explanation (MPE), is just one of such questions, but most relevant to our discussion.
The Viterbi dynamic programming algorithm solves this problem for chain networks in
O(L) time. Let the highest score of any subsequence from 1 to k > 1 ending with value y
k
be defined as
φ

k
(y
k
) = max
y
1..k−1

j
φ
j
(x, y
j

j
(x, y
j−1
, y
j
).
The algorithm computes the highest scores recursively:
φ

1
(y
1
) = φ
1
(x, y
1
), ∀y
1
∈ ¸
1
;
φ

k
(y
k
) = max
y
k−1
∈¸
k−1
φ

k−1
(y
k−1

j
(x, y
k

j
(x, y
k−1
, y
k
), 1 < k ≤ L, ∀y
k
∈ ¸
k
.
The highest scoring sequence has score max
y
L
φ

L
(y
L
). Using the arg max’s of the max’s in
the computation of φ

, we can back-trace the highest scoring sequence itself. We assume
that score ties are broken in a predetermined way, say according to some lexicographic
order of the symbols.
32 CHAPTER 3. STRUCTURED MODELS
Figure 3.3: Diamond Markov network (added triangulation edge is dashed).
In general Markov networks, MAP inference is NP-hard [Cowell et al., 1999]. How-
ever, there are several important subclasses of networks that allow polynomial time infer-
ence. The most important of these is the class of networks with low tree-width. We need the
concept of triangulation (or chordality) to formally define tree-width. Recall that a cycle
of length l in an undirected graph ( is a sequence of nodes (v
0
, v
1
, . . . , v
l
), distinct except
that v
0
= v
l
, which are connected by edges (v
i
, v
i+1
) ∈ (. A chord of this cycle is an edge
(v
i
, v
j
) ∈ ( between non-consecutive nodes.
Definition 3.3.2 (Triangulated graph) An undirected graph ( is triangulated if every one
of its cycles of length ≥ 4 possesses a chord.
Singly-connected graphs, like chains and trees, are triangulated since they contain no cy-
cles. The simplest untriangulated network is the diamond in Fig. 3.3. To triangulate it,
we can add the edge (Y
1
, Y
3
) or (Y
2
, Y
4
). In general, there are many possible sets of edges
that can be added to triangulate a graph. The inference procedure creates a tree of cliques
using the graph augmented by triangulation. The critical property of a triangulation for the
inference procedure is the size of the largest clique.
Definition 3.3.3 (Tree-width of a graph) The tree-width of a triangulated graph ( is the
size of its largest clique minus 1. The tree-width of an untriangulated graph ( is the
minimum tree-width of all triangulations of (.
3.3. MARKOV NETWORKS 33
The tree-width of a chain or a tree is 1 and the tree-width of Fig. 3.3 is 2. Finding the mini-
mum tree-width triangulation of a general graph is NP-hard, but good heuristic algorithms
exist [Cowell et al., 1999].
The inference procedure is based on a data structure called junction tree that can be
constructed for a triangulated graph. The junction tree is an alternative representation of
the same distribution that allows simple dynamic programming inference similar to the
Viterbi algorithm for chains.
Definition 3.3.4 (Junction tree) A junction tree T = (1, c) for a triangulated graph ( is
a tree in which the nodes are a subset of the cliques of the graph, 1 ⊆ ((() and the edges
c satisfy the running intersection property: for any two cliques c and c
t
, the variables in
the intersection c ∩ c
t
are contained in the clique of every node of the tree on the (unique)
path between c and c
t
.
Fig. 3.4 shows a junction tree for the diamond network. Each of the original clique poten-
tials must associated with exactly one node in the junction tree. For example, the potentials
for the ¦Y
1
, Y
3
, Y
4
¦ and ¦Y
1
, Y
3
, Y
4
¦ nodes are the product of the associated clique poten-
tials:
φ
134
(Y
1
, Y
3
, Y
4
) = φ
1
(Y
1

4
(Y
4

14
(Y
1
, Y
4

34
(Y
3
, Y
4
),
φ
123
(Y
1
, Y
2
, Y
3
) = φ
2
(Y
2

3
(Y
3

12
(Y
1
, Y
2

23
(Y
2
, Y
3
).
Algorithms for constructing junction trees from triangulated graphs are described in detail
in Cowell et al. [1999].
The Viterbi algorithm for junction trees picks an arbitrary root r for the tree T and
proceeds recursively from the leaves to compute the highest scoring subtree at a node by
combining the subtrees with highest score from its children. We denote the leaves of the
tree as Lv(T ) and the children of node c (relative to the root r) as Ch
r
(c):
φ

l
(y
l
) = φ
l
(x, y
l
), ∀l ∈ Lv(T ), ∀y
l
∈ ¸
l
;
φ

c
(y
c
) = φ
c
(x, y
c
)

c

∈Ch
r
(c)
max
y
c
∼y
c
φ

c
(y
c
), ∀c ∈ 1(T ) ¸ Lv(T ), ∀y
c
∈ ¸
c
,
34 CHAPTER 3. STRUCTURED MODELS
Figure 3.4: Diamond network junction tree. Each of the original potentials is associated
with a node in the tree.
where y
c
∼ y
c
denotes whether the partial assignment y
c
is consistent with the partial
assignment y
c
on the variables in the intersection of c and c
t
. The highest score is given by
max
y
r
φ

r
(y
r
). Using the arg max’s of the max’s in the computation of φ

, we can back-
trace the highest scoring assignment itself. Note that this algorithm is exponential in the
tree-width, the size of the largest clique. Similar type of computations using the junction
tree can be used to compute the partition function Z
w
(x) (by simply replacing max by

)
as well as marginal probabilities P(y
c
[x) for the cliques of the graph [Cowell et al., 1999].
3.3.3 Linear programming MAP inference
In this section, we present an alternative inference method based on linear programming.
Although solving the MAP inference using a general LP solver is less efficient than the
dynamic programming algorithms above, this formulation is crucial in viewing Markov
networks in a unified framework of the structured models we consider and to our develop-
ment of common estimation methods in later chapters. Let us begin with a linear integer
program to compute the optimal assignment y. We represent an assignment as a set binary
variables µ
c
(y
c
), one for each clique c and each value of the clique y
c
, that denotes whether
the assignment has that value, such that:
log

c
φ
c
(x, y
c
) =

c,y
c
µ
c
(y
c
) log φ
c
(x, y
c
).
3.3. MARKOV NETWORKS 35
0 0 0 0
0 0 1 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 1 0
0 0 0 0
0 0 0 0
0
1
0
0
0
1
0
0
0 0 1 0 0 0 1 0
Figure 3.5: Example of marginal agreement: row sums of µ
12
(y
1
, y
2
) agree with µ
1
(y
1
),
column sums agree with µ
2
(y
2
).
We call these variables marginals, as they correspond to the marginals of a distribution that
has all of its mass centered on the MAP instantiation (assuming it is unique). There are
several elementary constraints that such marginals satisfy. First, they must sum to one for
each clique. Second, the marginals for cliques that share variables are consistent. For any
clique c ∈ ( and a subclique s ⊂ c, the assignment of the subclique, µ
s
(y
s
), must be
consistent with the assignment of the clique, µ
c
(y
c
). Together, these constraints define a
linear integer program:
max

c,y
c
µ
c
(y
c
) log φ
c
(x, y
c
) (3.2)
s.t.

y
c
µ
c
(y
c
) = 1, ∀c ∈ (; µ
c
(y
c
) ∈ ¦0, 1¦, ∀c ∈ (, ∀y
c
;
µ
s
(y
s
) =

y

c
∼y
s
µ
c
(y
t
c
), ∀s, c ∈ (, s ⊂ c, ∀y
s
.
For example, in case the network is a chain or a tree, we will have node and edge marginals
that sum to 1 and agree with each other as in Fig. 3.5.
Clearly, for any assignment y
t
, we can define µ
c
(y
c
) variables that satisfy the above
constraints by setting µ
c
(y
c
) = 1I(y
t
c
= y
c
). We can also show that converse is true: any
valid setting of µ
c
(y
c
) corresponds to a valid assignment y. In fact,
36 CHAPTER 3. STRUCTURED MODELS
Lemma 3.3.5 For a triangulated network with unique MAP assignment, the integrality
constraint in the integer program in Eq. (3.2) can be relaxed and the resulting LP is guar-
anteed to have integer solutions.
A proof of this lemma appears in Wainwright et al. [2002]. Intuitively, the constraints force
the marginals µ
c
(y
c
) to correspond to some valid joint distribution over the assignments.
The optimal distribution with the respect to the objective puts all its mass on the MAP
assignment. If the MAP assignment is not unique, the value of the LP is the same as
the value of the integer program, and any linear combination of the MAP assignments
maximizes the LP.
In case the network is not triangulated, the set of marginals is not guaranteed to rep-
resent a valid distribution. Consider, for example, the diamond network in Fig. 3.3 with
binary variables, with the following edge marginals that are consistent with the constraints:
µ
12
(0, 0) = µ
12
(1, 1) = 0.5, µ
12
(1, 0) = µ
12
(0, 1) = 0;
µ
23
(0, 0) = µ
23
(1, 1) = 0.5, µ
23
(1, 0) = µ
23
(0, 1) = 0;
µ
34
(0, 0) = µ
34
(1, 1) = 0.5, µ
34
(1, 0) = µ
34
(0, 1) = 0;
µ
14
(0, 0) = µ
34
(1, 1) = 0, µ
14
(1, 0) = µ
14
(0, 1) = 0.5.
The corresponding node marginals must all be set to 0.5. Note that the edge marginals for
(1, 2), (2, 3), (3, 4) disallow any assignment other than 0000 or 1111, but the edge marginal
for (1, 4) disallows any assignment that has Y
1
= Y
4
. Hence this set of marginals dis-
allows all assignments. If we triangulate the graph and add the cliques ¦Y
1
, Y
2
, Y
3
¦ and
¦Y
1
, Y
3
, Y
4
¦ with their corresponding constraints, the above marginals will be disallowed.
In graphs where triangulation produces very large cliques, exact inference is intractable.
We can resort to the above LP without triangulation as an approximate inference procedure
(augmented with some procedure for rounding possibly fractional solutions). In Ch. 7, we
discuss another subclass of networks where MAP inference using LPs is tractable for any
network topology, but with a restricted type of potentials.
3.4. CONTEXT FREE GRAMMARS 37
Figure 3.6: Example parse tree from Penn Treebank [Marcus et al., 1993].
3.4 Context free grammars
Context-free grammars are one of the primary formalisms for capturing the recursive struc-
ture of syntactic constructions [Manning & Sch¨ utze, 1999]. For example, Fig. 3.6 shows
a parse tree for the sentence The screen was a sea of red. This tree is from the Penn Tree-
bank [Marcus et al., 1993], a primary linguistic resource for expert-annotated English text.
The non-terminal symbols (labels of internal nodes) correspond to syntactic categories such
as noun phrase (NP), verbal phrase (VP) or prepositional phrase (PP) and part-of-speech
tags like nouns (NN), verbs (VBD), determiners (DT) and prepositions (IN). The terminal
symbols (leaves) are the words of the sentence.
For clarity of presentation, we restrict our grammars to be in Chomsky normal form
2
(CNF),
where all rules in the grammar are of the form: A →B C and A →D, where A, B and C
are non-terminal symbols, and D is a terminal symbol.
Definition 3.4.1 (CFG) A CFG ( consists of:
◦ A set of non-terminal symbols, ^
◦ A designated set of start symbols, ^
S
⊆ ^
2
Any CFG can be represented by another CFG in CNF that generates the same set of sentences.
38 CHAPTER 3. STRUCTURED MODELS
◦ A set of terminal symbols, T
◦ A set of productions, T = ¦T
B
, T
U
¦, divided into
Binary productions, T
B
= ¦A →B C : A, B, C ∈ ^¦ and
Unary productions, T
U
= ¦A →D : A ∈ ^, D ∈ T ¦.
Consider a very simple grammar:
◦ ^ = ¦S, NP, VP, PP, NN, VBD, DT, IN¦
◦ ^
S
= ¦S¦
◦ T = ¦The, the, cat, dog, tree, saw, from¦
◦ T
B
= ¦S → NP VP, NP → DT NN, NP → NP PP, VP → VBD NP,
VP →VP PP, PP →IN NP¦.
◦ T
U
= ¦DT → The, DT → the, NN → cat, NN → dog, NN → tree, VBD → saw,
IN → from¦
A grammar generates a sentence by starting with a symbol in ^
S
and applying the
productions in T to rewrite nonterminal symbols. For example, we can generate The cat
saw the dog by starting with S → NP VP, rewriting the NP as NP → DT NN with DT →
The and NN → cat, then rewriting the VP as VP → VBD NP with VBD → saw, again
using NP → DT NN, but now with DT → the and NN → dog. We can represent such
derivations using trees like in Fig. 3.6 or (more compactly) using bracketed expressions
like the one below:
[[The
DT
cat
NN
]
NP
[saw
VBD
[the
DT
dog
NN
]
NP
]
VP
]
S
.
The simple grammar above can generate sentences of arbitrary length, since it has sev-
eral recursive productions. It can also generate the same sentence several ways. In general,
there are exponentially many parse trees that produce a sentence of length l. Consider the
sentence The cat saw the dog from the tree. The likely analysis of the sentence is that
the cat, sitting in the tree, saw the dog. An unlikely but possible alternative is that the cat
3.4. CONTEXT FREE GRAMMARS 39
actually saw the dog who lived near the tree or was tied to it in the past. Our grammar
allows both interpretations, with the difference being in the analysis of the top-level VP:
[saw
VBD
[the
DT
dog
NN
]
NP
]
VP
[[from
IN
[the
DT
tree
NN
]
NP
]
PP
,
saw
VBD
[[the
DT
dog
NN
]
NP
[from
IN
[the
DT
tree
NN
]
NP
]
PP
]
NP
.
This kind of ambiguity, called prepositional attachment, is very common in many re-
alistic grammars. A standard approach to resolving ambiguity is to use a PCFG to define
a joint probability distribution over the space of parse trees ¸ and sentences A. Standard
PCFG parsers use a Viterbi-style algorithm to compute arg max
y
P(x, y) as the most likely
parse tree for a sentence x. The distribution P(x, y) is defined by assigning a probability
to each production and making sure that the sum of probabilities of all productions starting
with a each symbol is 1:

B,C:A→B C∈1
B
P(A →B C) = 1,

D:A→D∈1
U
P(A → D ) = 1, ∀A ∈ ^.
We also need to assign a probability to the different starting symbols P(A) ∈ ^
S
such that

A∈A
S
P(A) = 1. The probability of a tree is simply the product of probabilities of the
productions used in the tree (times the probability of the starting symbol). Hence the log-
probability of a tree is a sum of the log-probabilities of its productions. By letting our basis
functions f (x, y) consist of the counts of the productions and w be their log-probabilities,
we can cast PCFG as a structured linear model (in log space). In Ch. 9, we will show how
to represent a parse tree as an assignment of variables Y with appropriate constraints to
express PCFGS (and more generally weighted CFGs) in the form of Eq. (3.1) as
h
w
(x) = arg max
y : g(x,y)≤0
w
¯
f (x, y),
and describe the associated algorithm to compute the highest scoring parse tree y given a
sentence x.
40 CHAPTER 3. STRUCTURED MODELS
3.5 Combinatorial problems
Many important computational tasks are formulated as combinatorial optimization prob-
lems such as the maximum weight bipartite and perfect matching, spanning tree, graph-cut,
edge-cover, bin-packing, and many others [Lawler, 1976; Papadimitriou & Steiglitz, 1982;
Cormen et al., 2001]. These problems arise in applications such as resource allocation,
job assignment, routing, scheduling, network design and many more. In some domains,
the weights of the objective function in the optimization problem are simple and natural
to define (for example, Euclidian distance or temporal latency), but in many others, con-
structing the weights is an important and labor-intensive design task. Treated abstractly, a
combinatorial space of structures, such as matchings or graph-cuts or trees), together with
a scoring scheme that assigns weights to candidate outputs is a kind of a model.
As a particularly simple and relevant example, consider modeling the task of assigning
reviewers to papers as a maximum weight bipartite matching problem, where the weights
represent the “expertise” of each reviewer for each paper. More specifically, suppose we
would like to have R reviewers per paper, and that each reviewer be assigned at most P pa-
pers. For each paper and reviewer, we have an a weight q
jk
indicating the qualification level
of reviewer j for evaluating paper k. Our objective is to find an assignment for reviewers
to papers that maximizes the total weight. We represent a matching with a set of binary
variables y
jk
that take the value 1 if reviewer j is assigned to paper k, and 0 otherwise. The
bipartite matching problem can be solved using a combinatorial algorithm or the following
linear program:
max

j,k
µ
jk
q
jk
(3.3)
s.t.

j
µ
jk
= R,

k
µ
jk
≤ P, 0 ≤ µ
jk
≤ 1.
This LP is guaranteed to produce integer solutions (as long as P and R are integers) for
any weights q(y) [Nemhauser & Wolsey, 1999].
The quality of the solution found depends critically on the choice of weights that de-
fine the objective. A simple scheme could measure the “expertise” as the percent of word
3.5. COMBINATORIAL PROBLEMS 41
overlap in the reviewer’s home page and the paper’s abstract. However, we would want to
weight certain words much more (words that are relevant to the subject and infrequent).
Constructing and tuning the weights for a problem is a difficult and time-consuming pro-
cess, just as it is for Markov networks for handwriting recognition.
As usual, we will represent the objective q(y) as a weighted combination of a set of
basis functions w
¯
f (x, y). Let x
jk
denote the intersection of the set of words occurring in
webpage(j)∩abstract(k), the web page of a reviewer j and the abstract of the paper k. We
can define f
d
(x, y) =

jk
y
jk
1I(word
d
∈ x
jk
), the number of times word d was in both
the web page of a reviewer and the abstract of the paper that were matched in y. Then the
score q
jk
is simply q
jk
=

d
w
d
1I(word
d
∈ x
jk
), a weighted combination of overlapping
words. In the next chapter we will show how to learn the parameters w in much the same
way we learn the parameters w of a Markov network.
The space of bipartite matchings illustrates an important property of many structured
spaces: the maximization problem arg max
y∈¸
w
¯
f (x, y) is easier than the normalization
problem

y∈¸
exp¦w
¯
f (x, y)¦. The maximum weight bipartite matching can be found
in polynomial (cubic) time in the number of nodes in the graph using a combinatorial algo-
rithm. However, even simply counting the number of matchings is #P-complete [Valiant,
1979; Garey & Johnson, 1979]. Note that counting is easier than normalization, which is
essentially weighted counting. This fact makes a probabilistic interpretation of the model as
a distribution over matchings intractable to compute. Similarly, exact maximum likelihood
estimation is intractable, since it requires computing the normalization.
Chapter 4
Structured maximum margin estimation
In the previous chapter, we described several important types of structured models of the
form:
h
w
(x) = arg max
y: g(x,y)≤0
w
¯
f (x, y), (4.1)
where we assume that the optimization problem max
y : g(x,y)≤0
w
¯
f (x, y) can be solved
or approximated by a compact convex optimization problem for some convex subset of
parameters w ∈ ¼. A compact problem formulation is polynomial in the description
length of the objective and the constraints.
Given a sample S = ¦(x
(i)
, y
(i)

m
i=1
, we develop methods for finding parameters w
such that:
arg max
y∈¸
(i)
w
¯
f (x
(i)
, y) ≈ y
(i)
, ∀i,
where ¸
(i)
= ¦y : g(x
(i)
, y) ≤ 0¦. In this chapter, we describe at an abstract level two
general approaches to structured estimation that we apply in the rest of the thesis. Both of
these approaches define a convex optimization problem for finding such parameters w.
There are several reasons to derive compact convex formulations. First and foremost,
we can find globally optimal parameters (with fixed precision) in polynomial time. Sec-
ond, we can use standard optimization software to solve the problem. Although special-
purpose algorithms that exploit the structure of a particular problem are often much faster
42
4.1. MAX-MARGIN ESTIMATION 43
(see Ch. 6), the availability of off-the-shelf software is very important for quick develop-
ment and testing of such models. Third, we can analyze the generalization performance of
the framework without worrying about the actual algorithms used to carry out the optimiza-
tion and the associated woes of intractable optimization problems: local minima, greedy
and heuristic methods, etc.
Our framework applies not only to the standard models typically estimated by prob-
abilistic methods, such as Markov networks and context-free grammars, but also to a
wide range of “unconventional” predictive models. Such models include graph cuts and
weighted matchings, where maximum likelihood estimation is intractable. We provide ex-
act maximum margin solutions for several of these problems (Ch. 7 and Ch. 10).
In prediction problems where the maximization in Eq. (4.1) is intractable, we consider
convex programs that provide only an upper or lower bound on the true solution. We
discuss how to use these approximate solutions for approximate learning of parameters.
4.1 Max-margin estimation
As in the univariate prediction, we measure the error of approximation using a loss func-
tion . In structured problems, where we are jointly predicting multiple variables, the loss
is often not just the simple 0-1 loss or squared error. For structured classification, a natural
loss function is a kind of Hamming distance between y
(i)
and h(x
(i)
): the number of vari-
ables predicted incorrectly. We will explore these and more general loss functions in the
following chapters.
4.1.1 Min-max formulation
Throughout, we will adopt the hinge upper bound
i
(h(x
(i)
)) on the loss function for struc-
tured classification inspired by max-margin criterion:

i
(h(x
(i)
)) = max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)] −w
¯
f
i
(y
(i)
) ≥
i
(h(x
(i)
)),
44 CHAPTER 4. STRUCTURED MAXIMUM MARGIN ESTIMATION
where as before,
i
(h(x
(i)
)) = (x
(i)
, y
(i)
, h(x
(i)
)),
i
(h(x
(i)
)) = (x
(i)
, y
(i)
, h(x
(i)
)), and
f
i
(y) = f (x
(i)
, y). With this upper bound, the min-max formulation for structured classifi-
cation problem is analogous to multi-class SVM formulation in Eq. (2.9) and Eq. (2.10):
min
1
2
[[w[[
2
+ C

i
ξ
i
(4.2)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)], ∀i.
The above formulation is a convex quadratic program in w, since max
y∈¸
(i) [w
¯
f
i
(y) +

i
(y)] is convex in w (maximum of affine functions is a convex function). For brevity, we
did not explicitly include the constraint that the parameters are in some legal convex set
(w ∈ ¼, most often IR
n
), but assume this throughout this chapter.
The problem with Eq. (4.2) is that the constraints have a very unwieldy form. An-
other way to express this problem is using

i

(i)
[ linear constraints, which is generally
exponential in L
i
, the number of variables in y
i
.
min
1
2
[[w[[
2
+ C

i
ξ
i
(4.3)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ w
¯
f
i
(y) +
i
(y), ∀i, ∀y ∈ ¸
(i)
.
This form reveals the “maximum margin” nature of the formulation. We can interpret
1
[[w[[
w
¯
[f
i
(y
(i)
) − f
i
(y)] as the margin of y
(i)
over another y ∈ ¸
(i)
. Assuming ξ
i
are all
zero (say because C is very large), the constraints enforce
w
¯
f
i
(y
(i)
) −w
¯
f
i
(y) ≥
i
(y),
so minimizing [[w[[ maximizes the smallest such margin, scaled by the loss
i
(y). The
slack variables ξ
i
allow for violations of the constraints at a cost Cξ
i
. If the loss function is
not uniform over all the mistakes y ,= y
(i)
, then the constraints make costly mistakes (those
with high
i
(y)) less likely. In Ch. 5 we analyze the effect of non-uniform loss function
(Hamming distance type loss) on generalization, and show a strong connection between the
loss-scaled margin and expected risk of the learned model.
The formulation in Eq. (4.3) is a standard QP with linear constraints, but its exponential
4.1. MAX-MARGIN ESTIMATION 45
size is in general prohibitive. We now return to Eq. (4.2) and transform it to a a more man-
ageable problem. The key to solving Eq. (4.2) efficiently is the loss-augmented inference
max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)]. (4.4)
Even if max
y∈¸
(i) w
¯
f
i
(y) can be solved in polynomial time using convex optimization, the
form of the loss term
i
(y) is crucial for the loss-augmented inference to remain tractable.
The range of tractable losses will depend strongly on the problem itself (f and ¸). Even
within the range of tractable losses, some are more efficiently computable than others. A
large part of the development of structured estimation methods in the following chapters
is identifying appropriate loss functions for the application and designing convex formula-
tions for the loss-augmented inference.
Assume that we find such a formulation in terms of a set of variables µ
i
, with a concave
(in µ
i
) objective
¯
f
i
(w, µ
i
) and subject to convex constraints ¯ g
i

i
):
max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)] = max
µ
i
: g
i

i
)≤0
¯
f
i
(w, µ
i
). (4.5)
We call such formulation compact if the number of variables µ
i
and constraints ¯ g
i

i
) is
polynomial in L
i
, the number of variables in y
(i)
.
Note that max
µ
i
: g
i

i
)≤0
¯
f
i
(w, µ
i
) must be convex in w, since Eq. (4.4) is. Likewise,
we can assume that it is feasible and bounded if Eq. (4.4) is. In the next section, we de-
velop a max-margin formulation that uses Lagrangian duality (see [Boyd & Vandenberghe,
2004] for an excellent review) to define a joint, compact convex problem for estimating the
parameters w.
To make the symbols concrete, consider the example of the reviewer-assignment prob-
lem we discussed in the previous chapter: we would like a bipartite matching with R re-
viewers per paper and at most P papers per reviewer. Each training sample i consists of a
matching of N
(i)
p
papers and N
(i)
r
reviewers from some previous year. Let x
jk
denote the
intersection of the set of words occurring in the web page of a reviewer j and the abstract of
the paper k. Let y
jk
indicate whether reviewer j is matched to the paper k. We can define
a basis function f
d
(x
jk
, y
jk
) = y
jk
1I(word
d
∈ x
jk
), which indicates whether the word d is
46 CHAPTER 4. STRUCTURED MAXIMUM MARGIN ESTIMATION
in both the web page of a reviewer and the abstract of the paper that are matched in y. We
abbreviate the vector of all the basis functions for each edge jk as y
jk
f
(i)
jk
= f (x
(i)
jk
, y
jk
).
We assume that the loss function decomposes over the variables y
ij
. For example, the
Hamming loss simply counts the number of different edges in the matchings y and y
(i)
:

H
i
(y) =

jk

0/1
i,jk
(y
jk
) =

jk
1I(y
jk
,= y
(i)
jk
) = RN
(i)
p

jk
y
jk
y
(i)
jk
.
The last equality follows from the fact that any valid matching for example i has R review-
ers for N
(i)
p
papers, hence RN
(i)
p

jk
y
jk
y
(i)
jk
represents exactly the number of edges that
are different between y and y
(i)
. Combining the two pieces, we have
w
¯
f (x
(i)
, y) =

jk
[w
¯
f (x
jk
, y
jk
) +
0/1
i,jk
(y
jk
)] = RN
(i)
p
+

jk
y
jk
[w
¯
f
jk
−y
(i)
jk
].
The loss-augmented inference problem can be then written as an LP in µ
i
similar
to Eq. (3.3) (without the constant term RN
(i)
p
):
max

jk
µ
i,jk
[w
¯
f
jk
−y
(i)
jk
]
s.t.

j
µ
i,jk
= R,

k
µ
i,jk
≤ P, 0 ≤ µ
i,jk
≤ 1.
In terms of Eq. (4.5),
¯
f
i
and ¯ g
i
are affine in µ
i
:
¯
f
i
(w, µ
i
) = RN
(i)
p
+

i,j
µ
i,jk
[w
¯
f
jk
−y
(i)
jk
]
and ¯ g
i

i
) ≤ 0 ⇔

j
µ
i,jk
= R,

k
µ
i,jk
≤ P, 0 ≤ µ
i,jk
≤ 1.
In general, when we can express max
y∈¸
(i) w
¯
f (x
(i)
, y) as an LP and we use a loss
function this is linear in the number of mistakes, we have a linear program of this form for
the loss-augmented inference:
d
i
+ max (F
i
w+c
i
)
¯
µ
i
s.t. A
i
µ
i
≤ b
i
, µ
i
≥ 0, (4.6)
for appropriately defined d
i
, F
i
, c
i
, A
i
, b
i
, which depend only on x
(i)
, y
(i)
, f (x, y) and
g(x, y). Note that the dependence on w is linear and only in the objective of the LP. If
this LP is compact (the number of variables and constraints is polynomial in the number of
4.1. MAX-MARGIN ESTIMATION 47
label variables), then we can use it to solve the max-margin estimation problem efficiently
by using convex duality.
The Lagrangian associated with Eq. (4.4) is given by
L
i,w

i
, λ
i
) =
¯
f
i
(w, µ
i
) −λ
¯
i
¯ g
i

i
), (4.7)
where λ
i
≥ 0 is a vector of Lagrange multipliers, one for each constraint function in
¯ g
i

i
). Since we assume that
¯
f
i
(w, µ
i
) is concave in µ
i
and bounded on the non-empty set
µ
i
: ¯ g
i

i
) ≤ 0, we have strong duality:
max
µ
i
: g
i

i
)≤0
¯
f
i
(w, µ
i
) = min
λ
i
≥0
max
µ
i
L
i,w

i
, λ
i
).
For many forms of
¯
f and ¯ g, we can write the Lagrangian dual min
λ
i
≥0
max
µ
i
L
i,w

i
, λ
i
)
explicitly as:
min h
i
(w, λ
i
) (4.8)
s.t. q
i
(w, λ
i
) ≤ 0,
where h
i
(w, λ
i
) and q
i
(w, λ
i
) are convex in both w and λ
i
. (We folded λ
i
≥ 0 into
q
i
(w, λ
i
) for brevity.) Since the original problem had polynomial size, the dual is polyno-
mial size as well. For example, the dual of the LP in Eq. (4.6) is
d
i
+ min b
¯
i
λ
i
s.t. A
¯
i
λ
i
≥ F
i
w+c
i
, λ
i
≥ 0, (4.9)
where h
i
(w, λ
i
) = d
i
+b
¯
i
λ
i
and q
i
(w, λ
i
) ≤ 0 is ¦F
i
w+c
i
−A
¯
i
λ
i
≤ 0, −λ
i
≤ 0¦.
Plugging Eq. (4.8) into Eq. (4.2), we get
min
1
2
[[w[[
2
+ C

i
ξ
i
(4.10)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ min
q
i
(w,λ
i
)≤0
h
i
(w, λ
i
), ∀i.
Moreover, we can combine the minimization over λ with minimization over ¦w, ξ¦. The
48 CHAPTER 4. STRUCTURED MAXIMUM MARGIN ESTIMATION
reason for this is that if the right hand side is not at the minimum, the constraint is tighter
than necessary, leading to a suboptimal solution w. Optimizing jointly over λ as well will
produce a solution to ¦w, ξ¦ that is optimal.
min
1
2
[[w[[
2
+ C

i
ξ
i
(4.11)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ h
i
(w, λ
i
), ∀i;
q
i
(w, λ
i
) ≤ 0, ∀i.
Hence we have a joint and compact convex optimization program for estimating w.
The exact form of this program depends strongly on
¯
f and ¯ g. For our LP-based example,
we have a QP with linear constraints:
min
1
2
[[w[[
2
+ C

i
ξ
i
(4.12)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ d
i
+b
¯
i
λ
i
, ∀i;
A
¯
i
λ
i
≥ F
i
w+c
i
, ∀i;
λ
i
≥ 0, ∀i.
4.1.2 Certificate formulation
In the previous section, we assumed a compact convex formulation of the loss-augmented
max in Eq. (4.4). There are several important combinatorial problems which allow poly-
nomial time solution yet do not have a compact convex optimization formulation. For
example, maximum weight perfect (non-bipartite) matching and spanning tree problems
can be expressed as linear programs with exponentially many constraints, but no polyno-
mial formulation is known [Bertsimas & Tsitsiklis, 1997; Schrijver, 2003]. Both of these
problems, however, can be solved in polynomial time using combinatorial algorithms. In
some cases, though, we can find a compact certificate of optimality that guarantees that
y
(i)
= arg max
y
[w
¯
f
i
(y) +
i
(y)] without expressing loss-augmented inference as a com-
pact convex program. Intuitively, just verifying that a given assignment is optimal is some-
times easier than actually finding it.
4.1. MAX-MARGIN ESTIMATION 49
Consider the maximum weight spanning tree problem. A basic property of a span-
ning tree is that cutting any edge (j, k) in the tree creates two disconnected sets of nodes
(1
j
[jk], 1
k
[jk]), where j ∈ 1
j
[jk] and k ∈ 1
k
[jk]. A spanning tree is optimal with respect
to a set of edge weights if and only if for every edge (j, k) in the tree connecting 1
j
[jk] and
1
k
[jk], the weight of (j, k) is larger than (or equal to) the weight of any other edge (j
t
, k
t
)
in the graph with j
t
∈ 1
j
[jk], k
t
∈ 1
k
[jk] [Cormen et al., 2001]. We discuss the conditions
for optimality of perfect matchings in Ch. 10. Suppose that we can find a compact convex
formulation of these conditions via a polynomial (in L
i
) set of functions q
i
(w, ν
i
), jointly
convex in w and auxiliary variables ν
i
:
∃ν
i
s.t. q
i
(w, ν
i
) ≤ 0 ⇔ w
¯
f
i
(y
(i)
) ≥ w
¯
f
i
(y) +
i
(y), ∀y ∈ ¸
(i)
.
Then the following joint convex program in wand ν computes the max-margin parameters:
min
1
2
[[w[[
2
(4.13)
s.t. q
i
(w, ν
i
) ≤ 0, ∀i.
Expressing the spanning tree optimality does not require additional variables ν
i
, but in
other problems, such as in perfect matching optimality in Ch. 10, such auxiliary variables
are needed. In the spanning tree problem, suppose y
jk
encodes whether edge (j, k) is in
the tree and the score of the edge is given by w
¯
f
i,jk
for some basis functions f (x
(i)
jk
, y
jk
).
We also assume that the loss function decomposes into a sum of losses over the edges, with
loss for each wrong edge given by
i,jk
. Then the optimality conditions are:
w
¯
f
i,jk
≥ w
¯
f
i,j

k
+
i,j

k
, ∀jk, j
t
k
t
s.t. y
(i)
jk
= 1, j
t
∈ 1
j
[jk], k
t
∈ 1
k
[jk].
For a full graph, we have
_
[1
(i)
[
3
) constraints for each example i, where [1
(i)
[ is the number
of nodes in the graph for example i.
Note that this formulation does not allow for violations of the margin constraints (it has
no slack variables ξ
i
). If the basis functions are not sufficiently rich to ensure that each y
(i)
is optimal, then Eq. (4.1.2) may be infeasible. Essentially, this formulation requires that
the upper bound on the empirical risk be zero, 1

S
[h
w
] = 0, and minimizes the complexity
50 CHAPTER 4. STRUCTURED MAXIMUM MARGIN ESTIMATION
of the hypothesis h
w
as measured by the norm of the weights.
If the problem is infeasible, the designer could add more basis functions f (x, y) that
take into account additional information about x. One could also add slack variables for
each example and each constraint that would allow violations of optimality conditions with
some penalty. However, these slack variables would not represent upper bounds on the loss
as they are in the min-max formulation, and therefore are less justified.
4.2 Approximations: upper and lower bounds
There are structured prediction tasks for which we might not be able to solve the estimation
problem exactly. Often, we cannot compute max
y∈¸
(i) [w
¯
f
i
(y) +
i
(y)] exactly or explic-
itly, but can only upper or lower bound it. Fig. 4.1 shows schematically how approximating
of the max subproblem reduces or extends the feasible space of w and ξ and leads to ap-
proximate solutions. The nature of these lower and upper bounds depends on the problem,
but we consider two general cases below.
4.2.1 Constraint generation
When neither compact maximization or optimality formulation is possible, but the max-
imization problem can be solved or approximated by a combinatorial algorithm, we can
resort to constraint generation or cutting plane methods. Consider Eq. (4.3), where we
have an exponential number of linear constraints, one for each i and y ∈ ¸
(i)
. Only a sub-
set of those constraints will be active at the optimal solution w. In fact, not more than the
number of parameters n plus the number of examples m can be active in general, since that
is the number of variables. If we can identify a small number of constraints that are critical
to the solution, we do not have to include all of them. Of course, identifying these con-
straints is in general as difficult as solving the problem, but a greedy approach of adding the
most violated constraints often achieves good approximate solutions after adding a small
(polynomial) number of constraints. If we continue adding constraints until there are no
more violated ones, the resulting solution is optimal.
We assume that we have an algorithm that produces y = arg max
y∈¸
(i) [w
¯
f
i
(y) +
4.2. APPROXIMATIONS: UPPER AND LOWER BOUNDS 51
Upper-bound
Exact
Lower-bound
×
+
Figure 4.1: Exact and approximate constraints on the max-margin quadratic program. The
solid red line represents the constraints imposed by the assignments y ∈ Y
(i)
, whereas the
dashed and dotted lines represent approximate constraints. The approximate constraints
may coincide with the exact constraints in some cases, and be more stringent or relaxed in
others. The parabolic contours represent the value of the objective function and ‘+’, ‘x’ and
‘o’ mark the different optima.

i
(y)]. The algorithm is described in Fig. 4.2. We maintain, for each example i, a small
but growing set of assignments
¯
¸
(i)
⊂ ¸
(i)
. At each iteration, we solve the problem with a
subset of constraints:
min
1
2
[[w[[
2
+ C

i
ξ
i
(4.14)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ w
¯
f
i
(y) +
i
(y), ∀i, ∀y ∈
¯
¸
(i)
.
The only difference between Eq. (4.3) and Eq. (4.14) is that ¸
(i)
has been replaced by
¯
¸
(i)
.
We then compute y = arg max
y∈¸
(i) [w
¯
f
i
(y) +
i
(y)] for each i and check whether the
constraint w
¯
f
i
(y
(i)
) + ξ
i
+ ≥ w
¯
f
i
(y) +
i
(y), is violated, where is a user defined
precision parameter. If it is violated, we set
¯
¸
(i)
=
¯
¸
(i)
∪ y. The algorithm terminates
when no constraints are violated. In Fig. 4.1, the lower-bound on the constraints provided
by
¯
¸
(i)
∪y keeps tightening with each iteration, terminating when the desired precision is
reached. We note that if the algorithm that produces y = arg max
y∈¸
(i) [w
¯
f
i
(y)+
i
(y)] is
52 CHAPTER 4. STRUCTURED MAXIMUM MARGIN ESTIMATION
Input: precision parameter .
1. Initialize:
¯
¸
(i)
= ¦¦, ∀ i.
2. Set violation = 0 and solve for w and ξ by optimizing
min
1
2
[[w[[
2
+ C

i
ξ
i
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ w
¯
f
i
(y) +
i
(y), ∀i, ∀y ∈
¯
¸
(i)
.
3. For each i,
Compute y = arg max
y∈¸
(i) [w
¯
f
i
(y) +
i
(y)],
if w
¯
f
i
(y
(i)
) + ξ
i
+ ≤ w
¯
f
i
(y) +
i
(y),
then set
¯
¸
(i)
=
¯
¸
(i)
∪ y and violation = 1
4. if violation = 1 goto 2.
Return w.
Figure 4.2: A constraint generation algorithm.
suboptimal, the approximation error of the solution we achieve might be much greater than
. The number of constraints that must be added before the algorithm terminates depends
on the precision and problem specific characteristics. See [Bertsimas & Tsitsiklis, 1997;
Boyd & Vandenberghe, 2004] for a more in-depth discussion of cutting planes methods.
This approach may also be computationally faster in providing a very good approximation
in practice if the explicit convex programming formulation is polynomial in size, but very
large, while the maximization algorithm is comparatively fast.
4.2.2 Constraint strengthening
In many problems, the maximization problem we are interested in may be very expensive
or intractable. For example, we consider MAP inference in large tree-width Markov net-
works in Ch. 8, multi-way cut in Ch. 7, graph-partitioning in Ch. 11. Many such problems
can be written as integer programs. Relaxations of such integer programs into LPs, QPs
or SDPs often provide excellent approximation algorithms [Hochbaum, 1997; Nemhauser
& Wolsey, 1999]. The relaxation usually defines a larger feasible space
¯
¸
(i)
⊃ ¸
(i)
over
4.3. RELATED WORK 53
which the maximization is done, where y ∈
¯
¸
(i)
may correspond to a “fractional” assign-
ment. For example, a solution to the MAP LP in Eq. (3.2) for an untriangulated network
may not correspond to any valid assignment. In such a case, the approximation is an over-
estimate of the constraints:
max
y∈

¸
(i)
[w
¯
f
i
(y) +
i
(y)] ≥ max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)].
Hence the constraint set is tightened with such invalid assignments. Fig. 4.1 shows how the
over-estimate reduces the feasible space of w and ξ.
Note that for every setting of the weights w that produces fractional solutions for the
relaxation, the approximate constraints are tightened because of the additional invalid as-
signments. In this case, the approximate MAP solution has higher value than any integer
solution, including the true assignment y
(i)
, thereby driving up the corresponding slack ξ
i
.
By contrast, for weights wfor which the MAP approximation is integer-valued, the margin
has the standard interpretation as the difference between the score of y
(i)
and the MAP y
(according to w). As the objective includes a penalty for the slack variable, intuitively,
minimizing the objective tends to drive the weights w away from the regions where the so-
lutions to the approximation are fractional. In essence, the estimation algorithm is finding
weights that are not necessarily optimal for an exact maximization algorithm, but (close to)
optimal for the particular approximate maximization algorithm used. In practice, we will
show experimentally that such approximations often work very well.
4.3 Related work
Our max-margin formulation is related to a body of work called inverse combinatorial and
convex optimization [Burton & Toint, 1992; Zhang & Ma, 1996; Ahuja & Orlin, 2001;
Heuberger, 2004]. An inverse optimization problem is defined by an instance of an opti-
mization problem max
y∈¸
w
¯
f (y), a set of nominal weights w
0
, and a target solution y
t
.
The goal is to find the weights w closest to the nominal w
0
in some norm, which make the
54 CHAPTER 4. STRUCTURED MAXIMUM MARGIN ESTIMATION
target solution optimal:
min [[w−w
0
[[
p
s.t. w
¯
f (y
t
) ≥ w
¯
f (y), ∀y ∈ ¸.
Most of the attention has been on L
1
and L

norms, but L
2
norm is also used.
The study of inverse problems began with geophysical scientists (see [Tarantola, 1987]
for in-depth discussion of a wide range of applications). Modeling a complex physical
system often involves a large number of parameters which scientists find hard or impossible
to set correctly. Provided educated guesses for the parameters w
0
and the behavior of the
system as a target, the inverse optimization problem attempts to match the behavior while
not perturbing the “guesstimate” too much.
Although there is a strong connection between inverse optimization problems and our
formulations, the goals are very different than ours. In our framework, we are learning
a parameterized objective function that depends on the input x and will generalize well
in prediction on new instances. Moreover, we do not assume as given a nominal set of
weights. Note that if we set w
0
= 0, then w = 0 is trivially the optimal solution. The
solution w depends critically on the choice of nominal weights, which is not appropriate in
the learning setting.
The inverse reinforcement learning problem [Ng & Russell, 2000; Abbeel & Ng, 2004]
is much closer to our setting. The goal is to learn a reward function that will cause a rational
agent to act similar to the observed behavior of an expert. A full description of the problem
is beyond our scope, but we briefly describe the Markov decision process (MDP) model
commonly used for sequential decision making problems where an agent interacts with its
environment. The environment is modeled as a system that can be in one of a set of discrete
states. At every time step, the agent chooses an action from a discrete set of actions and the
system transitions to a next state with a probability that depends on the current state and
the action taken. The agent collects a reward at each step, which generally depends on the
on the current and the next state and the action taken. A rational agent executes a policy
(essentially, a state to action mapping) that maximizes its expected reward. To map this
problem (approximately) to our setting, note that a policy roughly corresponds to the labels
4.4. CONCLUSION 55
y, the state sequence correspond to the input x and the reward for a state/action sequence is
assumed to be w
¯
f (x, y) for some basis functions w
¯
f (x, y). The goal is to learn wfroma
set of state/action sequences (x
(i)
, y
(i)
) of the expert such that the maximizing the expected
reward according to the system model makes the agent imitate the expert. This and related
problems are formulated as a convex program in Ng and Russell [2000] and Abbeel and
Ng [2004].
4.4 Conclusion
In this chapter, we presented two formulations of structured max-margin estimation that
define a compact convex optimization problem. The first formulation, min-max, relies on
the ability to express inference in the model as a compact convex optimization problem.
The second one, certificate, only requires expressing optimality of a given assignment ac-
cording to the model. Our framework applies to a wide range of prediction problems that
we explore in the rest of the thesis, including Markov networks, context free grammars, and
many combinatorial structures such as matchings and graph-cuts. The estimation problem
is tractable and exact whenever the prediction problem can be formulated as a compact
convex optimization problem or a polynomial time combinatorial algorithm with compact
convex optimality conditions. When the prediction problem is intractable or very expen-
sive to solve exactly, we resort to approximations that only provide upper/lower bounds
on the predictions. The estimated parameters are then approximate, but produce accurate
approximate prediction models in practice.
Because our approach only relies using the maximum in the model for prediction, and
does not require a normalized distribution P(y [ x) over all outputs, maximum margin
estimation can be tractable when maximum likelihood is not. For example, to learn a prob-
abilistic model P(y [ x) over bipartite matchings using maximum likelihood requires com-
puting the normalizing partition function, which is #P-complete [Valiant, 1979; Garey &
Johnson, 1979]. By contrast, maximum margin estimation can be formulated as a compact
QP with linear constraints. Similar results hold for non-bipartite matchings and min-cuts.
In models that are tractable for both maximum likelihood and maximum margin, (such
as low-treewidth Markov networks, context free grammars, many other problems in which
56 CHAPTER 4. STRUCTURED MAXIMUM MARGIN ESTIMATION
inference is solvable by dynamic programming), our approach has an additional advantage.
Because of the hinge-loss, the solutions to the estimation are relatively sparse in the dual
space (as in SVMs), which makes the use of kernels much more efficient. Maximum like-
lihood estimation with kernels results in models that are generally non-sparse and require
pruning or greedy support vector selection methods [Lafferty et al., 2004; Altun et al.,
2004].
The forthcoming formulations in the thesis follow the principles laid out in this chapter.
The range of applications of these principles is very broad and leads to estimation prob-
lems with very interesting structure in each particular problem, from Markov networks and
context-free grammars to graph cuts and perfect matchings.
Part II
Markov networks
57
Chapter 5
Markov networks
Markov networks are extensively used to model complex sequential, spatial, and relational
interactions in prediction problems arising in many fields. These problems involve labeling
a set of related objects that exhibit local consistency. In sequential labeling problems (such
as handwriting recognition), the labels (letters) of adjacent inputs (images) are highly corre-
lated. Sequential prediction problems arise in natural language processing (part-of-speech
tagging, speech recognition, information extraction [Manning & Sch¨ utze, 1999]), compu-
tational biology (gene finding, protein structure prediction, sequence alignment [Durbin
et al., 1998]), and many other fields. In image processing, neighboring pixels exhibit spa-
tial label coherence in denoising, segmentation and stereo correspondence [Besag, 1986;
Boykov et al., 1999a]. In hypertext or bibliographic classification, labels of linked and
co-cited documents tend to be similar [Chakrabarti et al., 1998; Taskar et al., 2002]. In
proteomic analysis, location and function of proteins that interact are often highly corre-
lated [Vazquez et al., 2003]. Markov networks compactly represent complex joint distribu-
tions of the label variables by modeling their local interactions. Such models are encoded
by a graph, whose nodes represent the different object labels, and whose edges represent
and quantify direct dependencies between them. For example, a Markov network for the
hypertext domain would include a node for each webpage, encoding its label, and an edge
between any pair of webpages whose labels are directly correlated (e.g., because one links
to the other).
We address the problem of max-margin estimation the parameters of Markov networks
58
5.1. MAXIMUM LIKELIHOOD ESTIMATION 59
for such structured classification problems. We show a compact convex formulation that
seamlessly integrates kernels with graphical models. We analyze the theoretical general-
ization properties of max-margin estimation and derive a novel margin-based bound for
structured classification.
We are given a labeled training sample S = ¦(x
(i)
, y
(i)

m
i=1
, drawn from a fixed dis-
tribution D over A ¸. We assume the structure of the network is given: we have a
mapping from an input x to the corresponding Markov network graph ((x) = ¦1, c¦
where the nodes 1 map to the variables in y. We abbreviate ((x
(i)
) as (
(i)
below. In hand-
writing recognition, this mapping depends on the segmentation algorithm that determines
how many letters the sample image contains and splits the image into individual images
for each letter. It also depends on the basis functions we use to model the dependencies of
the problem, for example, first-order Markov chain or a higher-order models. Note that the
topology and size of the graph (
(i)
, might be different for each example i. For instance, the
training sequences might have different lengths.
We focus on conditional Markov networks (or CRFs [Lafferty et al., 2001]), which
represent P(y [ x) instead of generative models P(x, y). The log-linear representation we
have described in Sec. 3.3.1 is defined via a vector of n basis functions f (x, y):
log P
w
(y [ x) = w
¯
f (x, y) −log Z
w
(x),
where Z
w
(x) =

y
exp¦w
¯
f (x, y)¦ and w ∈ IR
n
. Before we present the maximum
margin estimation, we review the standard maximum likelihood method.
5.1 Maximum likelihood estimation
The regularized maximum likelihood approach of learning the weights w of a Markov net-
work is similar to logistic regression we described in Sec. 2.2. The objective is to minimize
the training log-loss with an additional regularization term, usually the squared-norm of the
weights w [Lafferty et al., 2001]:
1
2
[[w[[
2
−C

i
log P
w
(y
(i)
[ x
(i)
) =
1
2
[[w[[
2
+ C

i
log Z
w
(x
(i)
) −w
¯
f
i
(y
(i)
),
60 CHAPTER 5. MARKOV NETWORKS
where f
i
(y) = f (x
(i)
, y).
This objective function is convex in the parameters w, so we have an unconstrained
convex optimization problem. The gradient with respect to w is given by:
w+ C

i
_
E
i,w
[f
i
(y)] −f
i
(y
(i)
)
¸
= w−C

i
E
i,w
[∆f
i
(y)],
where E
i,w
[f
i
(y)] =

y∈¸
f
i
(y)P
w
(y [ x
(i)
) is the expectation under the conditional dis-
tribution P
w
(y [ x
(i)
) and ∆f
i
(y) = f (x
(i)
, y
(i)
) −f (x
(i)
, y), as before. To compute the ex-
pectations, we can use inference in the Markov network to calculate marginals P
w
(y
c
[ x
(i)
)
for each clique c in the network Sec. 3.3.2. Since the basis functions decompose over the
cliques of the network, the expectation decomposes as well:
E
i,w
[f
i
(y)] =

c∈(
(i)

y
c
∈¸
(i)
c
f
i,c
(y
c
)P
w
(y
c
[ x
(i)
).
Second order methods for solving unconstrained convex optimization problems, such
as Newton’s method, require the second derivatives as well as the gradient. Let δf
i
(y) =
f
i
(y) − E
i,w
[f
i
(y)]. The Hessian of the objective depends on the covariances of the basis
functions:
I + C

i
E
i,w
_
δf
i
(y)δf
i
(y)
¯
¸
,
where I is a n n identity matrix. Computing the Hessian is more expensive than the
gradient, since we need to calculate joint marginals of every pair of cliques c and c
t
,
P
w
(y
c∪c
[ x
i
) as well as covariances of all basis functions, which is quadratic in the num-
ber of cliques and the number of functions. A standard approach is to use an approximate
second order method that does not need to compute the Hessian, but uses only the gradient
information [Nocedal & Wright, 1999; Boyd & Vandenberghe, 2004]. Conjugate Gradients
or L-BFGS methods have been shown to work very well on large estimation problems [Sha
& Pereira, 2003; Pinto et al., 2003], even with millions of parameters w.
5.2. MAXIMUM MARGIN ESTIMATION 61
5.2 Maximum margin estimation
For maximum-margin estimation, we begin with the min-max formulation from Sec. 4.1:
min
1
2
[[w[[
2
+ C

i
ξ
i
(5.1)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ max
y
[w
¯
f
i
(y) +
i
(y)], ∀i.
We know from Sec. 3.3.3 how to express max
y
w
¯
f
i
(y) as an LP, but the important differ-
ence is the loss function
i
. The simplest loss is the 0/1 loss
i
(y) ≡ 1I(y
(i)
,= y). In fact
this loss for sequence models was used by Collins [2001] and Altun et al. [2003]. However,
in structured problems, where we are predicting multiple labels, the loss is often not just
the simple 0/1 loss, but may depend on the number of labels and type of labels predicted
incorrectly or perhaps the number of cliques of labels predicted incorrectly. In general, we
assume that the loss, like the basis functions, decomposes over the cliques of labels.
Assumption 5.2.1 The loss function
i
(y) is decomposable:

i
(y) =

c∈((Ç
(i)
)
(x
(i)
c
, y
(i)
c
, y
c
) =

c∈((Ç
(i)
)

i,c
(y
c
).
We will focus on decomposable loss functions below. A natural choice that we use in our
experiments is the Hamming distance:

H
(x
(i)
, y
(i)
, y) =

v∈1
(i)
1I(y
(i)
v
,= y
v
).
With this assumption, we can express this inference problem for a triangulated graph
as a linear program for each example i as in Sec. 3.3.3:
max

c,y
c
µ
i,c
(y
c
)[w
¯
f
i,c
(y
c
) +
i,c
(y
c
)] (5.2)
s.t.

y
c
µ
i,c
(y
c
) = 1, ∀i, ∀c ∈ (
(i)
; µ
i,c
(y
c
) ≥ 0, ∀c ∈ (
(i)
, ∀y
c
;
µ
i,s
(y
s
) =

y

c
∼y
s
µ
i,c
(y
t
c
), ∀s, c ∈ (
(i)
, s ⊂ c, ∀y
s
,
62 CHAPTER 5. MARKOV NETWORKS
where (
(i)
= (((
(i)
) are the cliques of the Markov network for example i.
As we showed before, the constraints ensure that the µ
i
’s form a proper distribution. If
the most likely assignment is unique, then the distribution that maximizes the objective puts
all its weight on that assignment. (If the arg max is not unique, any convex combination of
the assignments is a valid solution). The dual of Eq. (5.2) is given by:
min

c
λ
i,c
(5.3)
s.t. λ
i,c
+

s⊃c
m
i,s,c
(y
c
) −

s⊂c, y

s
∼y
c
m
i,c,s
(y
t
s
) ≥ w
¯
f
i,c
(y
c
) +
i,c
(y
c
), ∀c ∈ (
(i)
, ∀y
c
.
In this dual, the λ
i,c
variables correspond to the normalization constraints, while m
i,c,s
(y
c
)
variables correspond to the agreement constraints in the primal in Eq. (5.2).
Plugging the dual into Eq. (5.1) for each example i and maximizing jointly over all the
variables (w, ξ, λ and m), we have:
min
1
2
[[w[[
2
+ C

i
ξ
i
(5.4)
s.t. w
¯
f
i
(y
(i)
) + ξ
i

i,c
λ
i,c
, ∀i;
λ
i,c
+

s⊃c
m
i,s,c
(y
c
) −

s⊂c, y

s
∼y
c
m
i,c,s
(y
t
s
) ≥ w
¯
f
i,c
(y
c
) +
i,c
(y
c
), ∀c ∈ (
(i)
, ∀y
c
.
In order to gain some intuition about this formulation, we make a change of variables from
λ
i,c
to ξ
i,c
:
λ
i,c
= w
¯
f
i,c
(y
(i)
c
) + ξ
i,c
, ∀i, ∀c ∈ (
(i)
.
The reason for naming the newvariables using the letter ξ will be clear in the following. For
readability, we also introduce variables that capture the effect of all the agreement variables
m:
M
i,c
(y
c
) =

s⊂c, y

s
∼y
c
m
i,c,s
(y
t
s
) −

s⊃c
m
i,s,c
(y
c
), ∀i, ∀c ∈ (
(i)
, ∀y
c
.
5.2. MAXIMUM MARGIN ESTIMATION 63
With these new variables, we have:
min
1
2
[[w[[
2
+ C

i
ξ
i
(5.5)
s.t. ξ
i

c
ξ
i,c
, ∀i;
w
¯
f
i,c
(y
(i)
c
) + ξ
i,c
≥ w
¯
f
i,c
(y
c
) +
i,c
(y
c
) + M
i,c
(y
c
), ∀i, ∀c ∈ (
(i)
, ∀y
c
;
M
i,c
(y
c
) =

s⊂c, y

s
∼y
c
m
i,c,s
(y
t
s
) −

s⊃c
m
i,s,c
(y
c
), ∀i, ∀c ∈ (
(i)
, ∀y
c
.
Note that ξ
i
=

c
ξ
i,c
at the optimum, since the slack variable ξ
i
only appears only in the
constraint ξ
i


c
ξ
i,c
and the objective minimizes Cξ
i
. Hence we can simply eliminate
this set of variables:
min
1
2
[[w[[
2
+ C

i,c
ξ
i,c
(5.6)
s.t. w
¯
f
i,c
(y
(i)
c
) + ξ
i,c
≥ w
¯
f
i,c
(y
c
) +
i,c
(y
c
) + M
i,c
(y
c
), ∀i, ∀c ∈ (
(i)
, ∀y
c
;
M
i,c
(y
c
) =

s⊂c, y

s
∼y
c
m
i,c,s
(y
t
s
) −

s⊃c
m
i,s,c
(y
c
), ∀i, ∀c ∈ (
(i)
, ∀y
c
.
Finally, we can write this in a form that resembles our original formulation Eq. (5.1), but
defined at a local level, for each clique:
min
1
2
[[w[[
2
+ C

i,c
ξ
i,c
(5.7)
s.t. w
¯
f
i,c
(y
(i)
c
) + ξ
i,c
≥ max
y
c
[w
¯
f
i,c
(y
c
) +
i,c
(y
c
) + M
i,c
(y
c
)], ∀i, ∀c ∈ (
(i)
;
M
i,c
(y
c
) =

s⊂c, y

s
∼y
c
m
i,c,s
(y
t
s
) −

s⊃c
m
i,s,c
(y
c
), ∀i, ∀c ∈ (
(i)
, ∀y
c
.
Note that without M
i,c
and m
i,c,s
variables, we essentially treat each clique as an indepen-
dent classification problem: for each clique we have a hinge upper-bound on the local loss,
or a margin requirement. The m
i,c,s
(y
s
) variables correspond to a certain kind of messages
between cliques that distribute “credit” to cliques to fulfill this margin requirement from
other cliques which have sufficient margin.
64 CHAPTER 5. MARKOV NETWORKS
Figure 5.1: First-order chain shown as a set of cliques (nodes and edges). Also shown are
the corresponding local slack variables ξ for each clique and messages m between cliques.
As an example, consider the first-order Markov chain in Fig. 5.1. The set of cliques
consists of the five nodes and the four edges. Suppose for the sake of this example that
our training data consists of only one training sample. The figure shows the local slack
variables ξ and messages m between cliques for this sample. For brevity of notion in this
example, we drop the dependence on the sample index i in the indexing of the variables
(we also used y
(∗)
j
instead of y
(i)
j
below). For concreteness, below we use the Hamming
loss
H
, which decomposes into local terms
j
(y
j
) = 1I(y
j
,= y
(∗)
j
) for each node and is
zero for the edges.
The constraints associated with the node cliques in this sequence are:
w
¯
f
1
(y
(∗)
1
) + ξ
1
≥ w
¯
f
1
(y
1
) + 1I(y
1
,= y
(∗)
1
) −m
1,12
(y
1
), ∀y
1
;
w
¯
f
2
(y
(∗)
2
) + ξ
2
≥ w
¯
f
2
(y
2
) + 1I(y
2
,= y
(∗)
2
) −m
2,12
(y
2
) −m
2,23
(y
2
), ∀y
2
;
w
¯
f
3
(y
(∗)
3
) + ξ
3
≥ w
¯
f
3
(y
3
) + 1I(y
3
,= y
(∗)
3
) −m
3,23
(y
3
) −m
3,34
(y
3
), ∀y
3
;
w
¯
f
4
(y
(∗)
4
) + ξ
4
≥ w
¯
f
4
(y
4
) + 1I(y
4
,= y
(∗)
4
) −m
4,34
(y
4
) −m
4,45
(y
4
), ∀y
4
;
w
¯
f
5
(y
(∗)
5
) + ξ
5
≥ w
¯
f
5
(y
5
) + 1I(y
5
,= y
(∗)
5
) −m
5,45
(y
5
), ∀y
5
.
5.3. M
3
N DUAL AND KERNELS 65
The edge constraints are:
w
¯
f
12
(y
(∗)
1
, y
(∗)
2
) + ξ
12
≥ w
¯
f
12
(y
1
, y
2
) + m
1,12
(y
1
) + m
2,12
(y
2
), ∀y
1
, y
2
;
w
¯
f
23
(y
(∗)
2
, y
(∗)
3
) + ξ
23
≥ w
¯
f
23
(y
2
, y
3
) + m
2,23
(y
2
) + m
3,23
(y
3
), ∀y
2
, y
3
;
w
¯
f
34
(y
(∗)
3
, y
(∗)
4
) + ξ
34
≥ w
¯
f
34
(y
3
, y
4
) + m
3,34
(y
3
) + m
4,34
(y
4
), ∀y
3
, y
4
;
w
¯
f
45
(y
(∗)
4
, y
(∗)
5
) + ξ
45
≥ w
¯
f
45
(y
4
, y
5
) + m
4,45
(y
4
) + m
5,45
(y
5
), ∀y
4
, y
5
.
5.3 M
3
N dual and kernels
In the previous section, we showed a derivation of a compact formulation based on LP
inference. In this section, we develop an alternative dual derivation that provides a very
interesting interpretation of the problem and is a departure for special-purpose algorithms
we develop. We begin with the formulation as in Eq. (4.3):
min
1
2
[[w[[
2
+ C

i
ξ
i
(5.8)
s.t. w
¯
∆f
i
(y) ≥
i
(y) −ξ
i
, ∀i, y,
where ∆f
i
(y) ≡ f (x
(i)
, y
(i)
) −f (x
(i)
, y). The dual is given by:
max

i,y
α
i
(y)
i
(y) −
1
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,y
α
i
(y)∆f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(5.9)
s.t.

y
α
i
(y) = C, ∀i; α
i
(y) ≥ 0, ∀i, y.
In the dual, the exponential number of α
i
(y) variables correspond to the exponential num-
ber of constraints in the primal. We make two small transformations to the dual that do not
change the problem: we normalize α’s by C (by letting α
i
(y) = Cα
t
i
(y)), so that they sum
66 CHAPTER 5. MARKOV NETWORKS
to 1 and divide the objective by C. The resulting dual is given by:
max

i,y
α
i
(y)
i
(y) −
1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,y
α
i
(y)∆f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(5.10)
s.t.

y
α
i
(y) = 1, ∀i; α
i
(y) ≥ 0, ∀i, y.
As in multi-class SVMs, the solution to the dual α gives the solution to the primal as a
weighted combination: w

= C

i,y
α

i
(y)∆f
i
(y).
Our main insight is that the variables α
i
(y) in the dual formulation Eq. (5.10) can be
interpreted as a kind of distribution over y, since they lie in the simplex

y
α
i
(y) = 1; α
i
(y) ≥ 0, ∀y.
This dual distribution does not represent the probability that the model assigns to an instan-
tiation, but the importance of the constraint associated with the instantiation to the solution.
The dual objective is a function of expectations of
i
(y) and ∆f
i
(y) with respect to α
i
(y).
Since
i
(y) =

c

i,c
(y
c
) and ∆f
i
(y) =

c
∆f
i,c
(y
c
) decompose over the cliques of the
Markov network, we only need clique marginals of the distribution α
i
(y) to compute their
expectations. We define the marginal dual variables as follows:
µ
i,c
(y
c
) =

y

∼y
c
α
i
(y
t
), ∀i, ∀c ∈ (
(i)
, ∀y
c
, (5.11)
where y
t
∼ y
c
denotes whether the partial assignment y
c
is consistent with the full assign-
ment y
t
. Note that the number of µ
i,c
(y
c
) variables is small (polynomial) compared to the
number of α
i
(y) variables (exponential) if the size of the largest clique is constant with
respect to the size of the network.
Now we can reformulate our entire QP (5.10) in terms of these marginal dual variables.
Consider, for example, the first term in the objective function (fixing a particular i):

y
α
i
(y)
i
(y) =

y
α
i
(y)

c

i,c
(y
c
) =

c,y
c

i,c
(y
c
)

y

∼y
c
α
i
(y
t
) =

c,y
c
µ
i,c
(y
c
)
i,c
(y
c
).
5.3. M
3
N DUAL AND KERNELS 67
The decomposition of the second term in the objective is analogous.

y
α
i
(y)∆f
i
(y) =

c,y
c
∆f
i,c
(y
c
)

y

∼y
c
α
i
(y
t
) =

c,y
c
µ
i,c
(y
c
)∆f
i,c
(y
c
).
Let us denote the the objective of Eq. (5.10) as Q(α). Note that it only depends on
α
i
(y) through its marginals µ
i,c
(y
c
), that is, Q(α) = Q
t
(/(α)), where / denotes the
marginalization operator defined by Eq. (5.11) . The domain of this operator, T[/], is
the product of simplices for all the m examples. What is its range, 1[/], the set of legal
marginals? Characterizing this set (also known as marginal polytope) compactly will allow
us to work in the space of µ’s:
max
α∈T[/]
Q(α) ⇔ max
µ∈1[/]
Q
t
(µ).
Hence we must ensure that µ
i
corresponds to some distribution α
i
, which is exactly
what the constraints in the LP for MAP inference enforce (see discussion of Lemma 3.3.5).
Therefore, when all (
(i)
are triangulated, the following structured dual QP has the same
primal solution (w

) as the original exponential dual QP in Eq. (5.10):
max

i,c,y
c
µ
i,c
(y
c
)
i,c
(y
c
) −
1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,c,y
c
µ
i,c
(y
c
)∆f
i,c
(y
c
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(5.12)
s.t.

y
c
µ
i,c
(y
c
) = 1, ∀i, ∀c ∈ (
(i)
; µ
i,c
(y
c
) ≥ 0, ∀i, ∀c ∈ (
(i)
, ∀y
c
;
µ
i,s
(y
s
) =

y

c
∼y
s
µ
i,c
(y
t
c
), ∀i, ∀s, c ∈ (
(i)
, s ⊂ c, ∀y
s
.
The solution to the structured dual µ

gives us the primal solution:
w

= C

i,c,y
c
µ

i,c
(y
c
)∆f
i,c
(y
c
).
In this structured dual, we only enforce that there exists an α
i
consistent with µ
i
, but do
not make a commitment about what it is. In general, the α distribution is not unique, but
there is a continuum of distributions consistent with a set of marginals. The objective of
68 CHAPTER 5. MARKOV NETWORKS
the QP Eq. (5.10) does not distinguish between these distributions, since it only depends on
their marginals. The maximum-entropy distribution α
i
consistent with a set of marginals
µ
i
, however, is unique for a triangulated model and can be computed using the junction tree
T
(i)
for the network [Cowell et al., 1999].
Specifically, associated with each edge (c, c
t
) in the tree T
(i)
is a set of variables called
the separator s = c ∩ c
t
. Note that each separator s and complement of a separator c ¸ s is
also a clique of the original graph, since it is a subclique of a larger clique. We denote the
set of separators as o
(i)
. Now we can define the maximum-entropy distribution α
i
(y) as
follows:
α
i
(y) =

c∈T
(i)
µ
i,c
(y
c
)

s∈S
(i)
µ
i,s
(y
s
)
. (5.13)
Again, by convention 0/0 ≡ 0.
Kernels
Note that the solution is a weighted combination of local basis functions and the objective
of Eq. (5.12) can be expressed in terms of dot products between local basis functions
∆f
i,c
(y
c
)
¯
∆f
j,¯ c
(y
¯ c
) = [f (x
(i)
c
, y
(i)
c
) −f (x
(i)
c
, y
c
)]
¯
[f (x
(j)
¯ c
, y
(j)
¯ c
) −f (x
(j)
¯ c
, y
¯ c
)].
Hence, we can locally kernelize our models and solve Eq. (5.12) efficiently. Kernels are
typically defined on the input, e.g. k(x
(i)
c
, x
(j)
¯ c
). In our handwriting example, we use a
polynomial kernel on the pixel values for the node cliques. We usually extend the kernel
over the input space to the joint input and output space by simply defining
f (x
c
, y
c
)
¯
f (x
¯ c
, y
¯ c
) ≡ 1I(y
c
= y
¯ c
)k(x
c
, x
¯ c
).
Of course, other definitions are possible and may be useful when the assignments in each
clique y
c
have interesting structure. In Sec. 6.2 we experiment with several kernels for
the handwriting example. As in SVMs, the solutions to the max-margin QP are typically
sparse in the µ variables. Hence, each log-potential in the network “remembers” only a
small proportion of the relevant training data inputs.
5.4. UNTRIANGULATED MODELS 69
Figure 5.2: Diamond Markov network (added triangulation edge is dashed and three-node
marginals are in dashed rectangles).
5.4 Untriangulated models
If the underlying Markov net is not chordal, we must address the problem by triangulating
the graph, that is, adding fill-in edges to ensure triangulation. For example, if our graph is
a 4-cycle Y
1
—Y
2
—Y
3
—Y
4
—Y
1
as in Fig. 5.2, we can triangulate the graph by adding an
arc Y
1
—Y
3
. This will introduce new cliques Y
1
, Y
2
, Y
3
and Y
1
, Y
3
, Y
4
and the corresponding
marginals, µ
123
(y
1
, y
2
, y
3
) and µ
134
(y
1
, y
3
, y
4
). We can then use this new graph to produce
the constraints on the marginals:

y
1
µ
123
(y
1
, y
2
, y
3
) = µ
23
(y
2
, y
3
), ∀y
2
, y
3
;

y
3
µ
123
(y
1
, y
2
, y
3
) = µ
12
(y
1
, y
2
), ∀y
1
, y
2
;

y
1
µ
134
(y
1
, y
3
, y
4
) = µ
34
(y
3
, y
4
), ∀y
3
, y
4
;

y
3
µ
134
(y
1
, y
3
, y
4
) = µ
13
(y
1
, y
3
), ∀y
1
, y
3
.
The new marginal variables appear only in the constraints; they do not add any new basis
functions nor change the objective function.
70 CHAPTER 5. MARKOV NETWORKS
In general, the number of constraints introduced is exponential in the number of vari-
ables in the new cliques — the tree-width of the graph. Unfortunately, even sparsely con-
nected networks, for example 2D grids often used in image analysis, have large tree-width.
However, we can still solve the QP in the structured primal Eq. (5.6) or the structured
dual Eq. (5.12) defined by an untriangulated graph. Such a formulation, which enforces
only local consistency of marginals, optimizes our objective only over a relaxation of the
marginal polytope. However, the learned parameters produce very accurate approximate
models in practice, as experiments in Ch. 8 demonstrate.
Note that we could also strengthen the untriangulated relaxation without introducing
an exponential number of constraints. For example, we can add positive semidefinite con-
straints on the marginals µ used by Wainwright and Jordan [2003], which tend to improve
the approximation of the marginal polytope. Although this and other more complex relax-
ations are a very interesting area of future development, they are often much more expen-
sive.
The approximate QP does not guarantee that the learned model using exact inference
minimizes the true objective: (upper-bound on) empirical risk plus regularization. But do
we really need these optimal parameters if we cannot perform exact inference? A more
useful goal is to make sure that training error is minimized using the approximate infer-
ence procedure via the untriangulated LP. We conjecture that the parameters learned by
the approximate QP in fact do that to some degree. For instance, consider the separable
case, where 100% accuracy is achievable on the training data by some parameter setting w
such that approximate inference (using the untriangulated LP) produces integral solutions.
Solving the problem as C → ∞ will find this solution even though it may not be optimal
(in terms of the norm of the w) using exact inference. For C in intermediate range, the
formulation trades off fractionality of the untriangulated LP solutions with complexity of
the weights [[w[[
2
.
5.5 Generalization bound
In this section, we show a generalization bound for the task of structured classification that
allows us to relate the error rate on the training set to the generalization error. To the best
5.5. GENERALIZATION BOUND 71
of our knowledge, this bound is the first to deal with structured error, such as the Hamming
distance. Our analysis of Hamming loss allows to prove a significantly stronger result than
previous bounds for the 0/1 loss, as we detail below.
Our goal in structured classification is often to minimize the number of misclassified
labels, or the Hamming distance between y and h(x). An appropriate error function is the
average per-label loss
L(w, x, y) =
1
L

H
(y, arg max
y

w
¯
f (x, y
t
)),
where Lis the number of label variables in y. As in other generalization bounds for margin-
based classifiers, we relate the generalization error to the margin of the classifier. Consider
an upper bound on the above loss:
L(w, x, y) ≤ L(w, x, y) = max
y

: w

f (y)≤w

f (y

)
1
L

H
(y, y
t
).
This upper bound is tight if y = arg max
y
w
¯
f (x, y
t
), Otherwise, it is adversarial: it
picks from all y
t
which are better (w
¯
f (y) ≤ w
¯
f (y
t
)), one that maximizes the Hamming
distance from y. We can now define a γ-margin per-label loss:
L(w, x, y) ≤ L(w, x, y) ≤ L
γ
(w, x, y) = max
y

: w

f (y)≤w

f (y

)+γ
H
(y,y

)
1
L

H
(y, y
t
).
This upper bound is even more adversarial: it is tight if y = arg max
y
[w
¯
f (x, y
t
) +

H
(y, y
t
)], otherwise, it picks from all y
t
which are better when helped by γ
H
(y, y
t
), one
that maximizes the Hamming distance from y. Note that the loss we minimize in the max-
margin formulation is very closely related (although not identical to) this upper bound.
We can now prove that the generalization accuracy of any hypothesis w is bounded by
its empirical γ-margin per-label loss, plus a term that grows inversely with the margin.To
state the bound, we need to define several other factors it depends upon. Let N
c
be the
maximum number of cliques in ((x), V
c
be the maximum number of values in a clique
[Y
c
[, q be the maximum number of cliques that have a variable in common, and R
c
be
an upper-bound on the 2-norm of clique basis functions. Consider a first-order sequence
72 CHAPTER 5. MARKOV NETWORKS
model as an example, with L as the maximumlength, and V the number of values a variable
takes. Then N
c
= 2L − 1 since we have L node cliques and L − 1 edge cliques; V
c
= V
2
because of the edge cliques; and q = 3 since nodes in the middle of the sequence participate
in 3 cliques: previous-current edge clique, node clique, and current-next edge clique.
Theorem 5.5.1 For the family of hypotheses parameterized by w, and any δ > 0, there
exists a constant K such that for any γ > 0 per-label margin, and m > 1 samples, the
expected per-label loss is bounded by:
E
D
[L(w, x, y)] ≤ E
S
[L
γ
(w, x, y)] +
¸
K
m
_
R
2
c
[[w[[
2
q
2
γ
2
[ln m + ln N
c
+ ln V
c
] + ln
1
δ
_
,
with probability at least 1 −δ.
Proof: See Appendix A.1 for the proof details and the exact value of the constant K.
The first term upper bounds the training error of w. Low loss E
S
[L
γ
(w, x, y)] at high
margin γ quantifies the confidence of the prediction model. The second term depends on
[[w[[/γ, which corresponds to the complexity of the classifier (normalized by the margin
level). Thus, the result provides a bound to the generalization error that trades off the
effective complexity of the hypothesis space with the training error.
The proof uses a covering number argument analogous to previous results in SVMs [Zhang,
2002]. However we propose a novel method for covering the space of structured prediction
models by using a cover of the individual clique basis function differences ∆f
i,c
(y
c
). This
new type of cover is polynomial in the number of cliques, yielding significant improve-
ments in the bound. Specifically, our bound has a logarithmic dependence on the number
of cliques (ln N
c
) and depends only on the 2-norm of the basis functions per-clique (R
c
).
This is a significant gain over the previous result of Collins [2001] for 0/1 loss, which has
linear dependence (inside the square root) on the number of nodes (L), and depends on
the joint 2-norm of all of the basis functions for an example (which is ∼ N
c
R
c
). Such a
result was, until now, an open problem for margin-based sequence classification [Collins,
2001]. Finally, for sequences, note that if
L
m
= O(1) (for example, in OCR, if the number
of instances is at least a constant times the length of a word), then our bound is independent
of the number of labels L.
5.6. RELATED WORK 73
5.6 Related work
The application of margin-based estimation methods to parsing and sequence modeling was
pioneered by Collins [2001] using the Voted-Perceptron algorithm [Freund & Schapire,
1998]. He provides generalization guarantees (for 0/1 loss) that hold for separable case and
depend on the number of mistakes the perceptron makes before convergence. Remarkably,
the bound does not explicitly depend on the length of the sequence, although undoubtedly
the number of mistakes does.
Collins [2004] also suggested an SVM-like formulation (with exponentially many con-
straints) and a constraint generation method for solving it. His generalization bound (for
0/1 loss) based on the SVM-like margin, however, has linear dependence (inside the square
root) on the number of nodes (L). It also depends on the joint 2-norm of all of the basis
functions for an example (which is ∼ N
c
R
c
). By considering the more natural Hamming
loss, we achieve a much tighter analysis.
Altun et al. [2003] have applied the exponential-size formulation with constraint gen-
eration we described in Sec. 4.2.1 to problems natural language processing. In a follow-up
paper, Tsochantaridis et al. [2004] show that only a polynomial number of constraints are
needed to be generated to guarantee a fixed level of precision of the solution. However,
the number of constraints in many important cases is several orders higher (in L) than in
the the approach we present. In addition, the corresponding problem needs to be resolved
(or at least approximately resolved) after each additional constraint is added, which is pro-
hibitively expensive for large number of examples and label variables.
The work of Guestrin et al. [2003] presents LP decompositions based on graphical
model structure for the value function approximation problem in factored MDPs (Markov
decision processes with structure). Describing the exact setting is beyond our scope, but it
suffices to say that our original decomposition of the max-margin QP was inspired by the
proposed technique to transform an exponential set of constraints into a polynomial one
using a triangulated graph.
There has been a recent explosion of work in maximum conditional likelihood estima-
tion of Markov networks. The work of Lafferty et al. [2001] has inspired many applications
in natural language, computational biology, computer vision and relational modeling [Sha
74 CHAPTER 5. MARKOV NETWORKS
& Pereira, 2003; Pinto et al., 2003; Kumar & Hebert, 2003; Sutton et al., 2004; Taskar
et al., 2002; Taskar et al., 2003b]. As in the case of logistic regression, maximum condi-
tional likelihood estimation for Markov networks can also be kernelized [Altun et al., 2004;
Lafferty et al., 2004]. However, the solutions are non-sparse and the proposed algorithms
are forced to use greedy selection of support vectors or heuristic pruning methods.
5.7 Conclusion
We use graph decomposition to derive an exact, compact, convex max-margin formulation
for Markov networks with sequence and other low-treewidth structure. Our formulation
avoids the exponential blow-up in the number of constraints in the max-margin QP that
plagued previous approaches. The seamless integration of kernels with graphical models
allows us to create very rich models that leverage the immense amount of research in kernel
design and graphical model decompositions. We also use approximate graph decomposi-
tion to derive a compact approximate formulation for Markov networks in which inference
is intractable.
We provide theoretical guarantees on the average per-label generalization error of our
models in terms of the training set margin. Our generalization bound significantly tightens
previous results of Collins [Collins, 2001] and suggests possibilities for analyzing per-label
generalization properties of graphical models.
In the next chapter, we present an efficient algorithm that exploits graphical model
inference and show experiments on a large handwriting recognition task that utilize the
powerful representational capability of kernels.
Chapter 6
M
3
N algorithms and experiments
Although the number of variables and constraints in the structured dual in Eq. (5.12) is
polynomial in the size of the data, unfortunately, for standard QP solvers, the problem is
often too large even for small training sets. Instead, we use a coordinate dual ascent method
analogous to the sequential minimal optimization (SMO) used for SVMs [Platt, 1999].
We apply our M
3
N framework and structured SMO algorithm to a handwriting recogni-
tion task. We show that our models significantly outperform other approaches by incorpo-
rating high-dimensional decision boundaries of polynomial kernels over character images
while capturing correlations between consecutive characters.
6.1 Solving the M
3
N QP
Let us begin by considering the primal and dual QPs for multi-class SVMs:
min
1
2
[[w[[
2
+ C

i
ξ
i
max

i,y
α
i
(y)
i
(y) −
1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,y
α
i
(y)∆f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
s.t. w

∆f
i
(y) ≥ (y) −ξ
i
, ∀i, y. s.t.

y
α
i
(y) = 1, ∀i; α
i
(y) ≥ 0, ∀i, y.
The KKT conditions [Bertsekas, 1999; Boyd & Vandenberghe, 2004] provide sufficient
and necessary criteria for optimality of a dual solution α. As we describe below, these
conditions have certain locality with respect to each example i, which allows us to perform
75
76 CHAPTER 6. M
3
N ALGORITHMS AND EXPERIMENTS
the search for optimal α by repeatedly considering one example at a time.
A feasible dual solution α and a primal solution defined by:
w = C

i,y
α
i
(y)∆f
i
(y) (6.1)
ξ
i
= max
y
[
i
(y) −w∆f
i
(y)] = max
y
[
i
(y) +w
¯
f
i
(y)] −w
¯
f
i
(y
(i)
),
are optimal if they satisfy the following two types of constraints:
α
i
(y) = 0 ⇒ w
¯
∆f
i
(y) >
i
(y) −ξ
i
; (KKT1)
α
i
(y) > 0 ⇒ w
¯
∆f
i
(y) =
i
(y) −ξ
i
. (KKT2),
We can express these conditions as
α
i
(y) = 0 ⇒ w
¯
f
i
(y) +
i
(y) < max
y

[w
¯
f
i
(y
t
) +
i
(y
t
)]; (KKT1)
α
i
(y) > 0 ⇒ w
¯
f
i
(y) +
i
(y) = max
y

[w
¯
f
i
(y
t
) +
i
(y
t
)]. (KKT2)
To simplify the notation, we define
v
i
(y) = w
¯
f
i
(y) +
i
(y); v
i
(y) = max
y

,=y
[w
¯
f
i
(y
t
) +
i
(y
t
)].
With these definitions, we have
α
i
(y) = 0 ⇒v
i
(y) < v
i
(y); (KKT1) α
i
(y) > 0 ⇒v
i
(y) ≥ v
i
(y); (KKT2).
In practice, however, we will enforce KKT conditions up to a given tolerance 0 < ¸1.
α
i
(y) = 0 ⇒v
i
(y) ≤ v
i
(y) + ; α
i
(y) > 0 ⇒v
i
(y) ≥ v
i
(y) −. (6.2)
Essentially, α
i
(y) can be zero only if v
i
(y) is at most larger than the all others. Con-
versely, α
i
(y) can be non-zero only if v
i
(y) is at most smaller than the all others.
Note that the normalization constraints on the dual variables α are local to each exam-
ple i. This allows us to perform dual block-coordinate ascent where a block corresponds to
6.1. SOLVING THE M
3
N QP 77
1. Initialize: α
i
(y) = 1I(y = y
(i)
), ∀ i, y.
2. Set violation = 0,
3. For each i,
4. If α
i
violates (KKT1) or (KKT2),
5. Set violation = 1,
6. Find feasible α
t
i
such that Q(α
t
i
, α
−i
) > Q(α
i
, α
−i
) and set α
i
= α
t
i
.
7. If violation = 1 goto 2.
Figure 6.1: Block-coordinate dual ascent.
the vector of dual variables α
i
for a single example i. The general form of block-coordinate
ascent algorithm as shown in Fig. 6.1 is essentially coordinate ascent on blocks α
i
, main-
taining the feasibility of the dual. When optimizing with respect to a single block i, the
objective function can be split into two terms:
Q(α) = Q(α
−i
) +Q(α
i
, α
−i
),
where α
−i
denotes all dual α
k
variables for k other than i. Only the second part of the
objective Q(α
i
, α
−i
) matters for optimizing with respect to α
i
. The algorithm starts with
a feasible dual solution α and improves the objective block-wise until all KKT condi-
tions are satisfied. Checking the constraints requires computing w and ξ from α according
to Eq. (6.1).
As long as the local ascent step over α
i
is guaranteed to improve the objective when
KKT conditions are violated, the algorithm will converge to the global maximum in a finite
number of steps (within the precision). This allows us to focus on efficient updates to a
single block of α
i
at a time.
Let α
t
i
(y) = α
i
(y) + λ(y). Note that

y
λ(y) = 0 and α
i
(y) + λ(y) ≥ 0 so that α
t
i
is
feasible. We can write the objective Q(α
−i
) +Q(α
t
i
, α
−i
) in terms of λ and α:

j,y
α
j
(y)
j
(y) +

y
λ(y)
i
(y) −
1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

y
λ(y)∆f
i
(y) +

j,y
α
j
(y)∆f
j
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
.
78 CHAPTER 6. M
3
N ALGORITHMS AND EXPERIMENTS
By dropping all terms that do not involve λ, and making the substitution
w = C

j,y
α
j
(y)∆f
j
(y), we get:

y
λ(y)
i
(y) −w
¯
_

y
λ(y)∆f
i
(y)
_

1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

y
λ(y)∆f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
.
Since

y
λ(y) = 0,

y
λ(y)∆f
i
(y) =

y
λ(y)f
i
(y
(i)
) −

y
λ(y)f
i
(y) = −

y
λ(y)f
i
(y).
Below we also make the substitution v
i
(y) = w
¯
f
i
(y) +
i
(y) to get the optimization
problem for λ:
max

y
λ(y)v
i
(y) −
1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

y
λ(y)f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
s.t.

y
λ(y) = 0; α
i
(y) + λ(y) ≥ 0, ∀y.
6.1.1 SMO
We do not need to solve the optimization subproblem above at each pass through the data.
All that is required is an ascent step, not a full optimization. Sequential Minimal Opti-
mization (SMO) approach takes an ascent step that modifies the least number of variables.
In our case, we have the simplex constraint, so we must change at least two variables in
order to respect the normalization constraint (by moving weight from one dual variable to
another). We address a strategy for selecting the two variables in the next section, but for
now assume we have picked λ(y
t
) and λ(y
tt
). Then we have δ = λ(y
t
) = −λ(y
tt
) in order
to sum to 1. The optimization problem becomes a single variable quadratic program in δ:
max [v
i
(y
t
) −v
i
(y
tt
)]δ −
1
2
C[[f
i
(y
t
) −f
i
(y
tt
)[[
2
δ
2
(6.3)
s.t. α
i
(y
t
) + δ ≥ 0; α
i
(y
tt
) −δ ≥ 0.
6.1. SOLVING THE M
3
N QP 79
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
−2
−1.5
−1
−0.5
0
0.5
Figure 6.2: Representative examples of the SMO subproblem. Horizonal axis represents δ
with two vertical lines depicting the upper and lower bounds c and d. Vertical axis repre-
sents the objective. Optimum either occurs at the maximum of the parabola if it is feasible
or the upper or lower bound otherwise.
With a = v
i
(y
t
) −v
i
(y
tt
), b = C[[f
i
(y
t
) −f
i
(y
tt
)[[
2
, c = −α
i
(y
t
), d = α
i
(y
tt
), we have:
max [aδ −
b
2
δ
2
] s.t. c ≤ δ ≤ d, (6.4)
where the optimum is achieved at the maximum of the parabola a/b if c ≤ a/b ≤ d or at
the boundary c or d (see Fig. 6.1.1). Hence the solution is given by simply clipping a/b:
δ

= max(c, min(d, a/b)).
The key advantage of SMO is the simplicity of this update. Computing the coefficients
involves dot products (or kernel evaluations) to compute w
¯
f
i
(y
t
) and w
¯
f
i
(y
tt
) as well as
(f
i
(y
t
) −f
i
(y
tt
))
¯
(f
i
(y
t
) −f
i
(y
tt
)).
6.1.2 Selecting SMO pairs
How do we actually select such a pair to guarantee that we make progress in optimizing
the objective? Note that at least one of the assignments y must violate (KKT1) or (KKT2),
80 CHAPTER 6. M
3
N ALGORITHMS AND EXPERIMENTS
1. Set violation = 0.
2. For each y,
3. KKT1: If α
i
(y) = 0 and v
i
(y) > v
i
(y) + ,
4. Set y
t
= y and violation = 1 and goto 7.
5. KKT2: If α
i
(y) > 0 and v
i
(y) < v
i
(y) −,
6. Set y
t
= y and violation = 2 and goto 7.
7. If violation > 0,
8. For each y ,= y
t
,
9. If violation = 1 and α
i
(y) > 0,
10. Set y
tt
= y and goto 13.
11. If violation = 2 and v
i
(y) > v
i
(y
t
),
12. Set y
tt
= y and goto 13.
13. Return y
t
and y
tt
.
Figure 6.3: SMO pair selection.
because otherwise α
i
is optimal with respect to the current α
−i
. The selection algorithm is
outlined in Fig. 6.3.
The first variable in the pair, y
t
, corresponds to a violated condition, while the second
variable, y
tt
, is chosen to guarantee that solving Eq. (6.3) will result in improving the ob-
jective. There are two cases, corresponding to violation of KKT1 and violation of KKT2.
Case KKT1. α
i
(y
t
) = 0 but v
i
(y
t
) > v
i
(y
t
) + . This is the case where i, y
t
is a not
support vector but should be. We would like to increase α
i
(y
t
), so we need α
i
(y
tt
) > 0
to borrow from. There will always be a such a y
tt
since

y
α
i
(y) = 1 and α
i
(y
t
) = 0.
Since v
i
(y
t
) > v
i
(y
t
) + , v
i
(y
t
) > v
i
(y
tt
) + , so the linear coefficient in Eq. (6.4) is
a = v
i
(y
t
) −v
i
(y
tt
) > . Hence the unconstrained maximum is positive a/b > 0. Since the
upper-bound d = α
i
(y
tt
) > 0, we have enough freedom to improve the objective.
Case KKT2. α
i
(y
t
) > 0 but v
i
(y
t
) < v
i
(y
t
) −. This is the case where i, y
t
is a support
vector but should not be. We would like to decrease α
i
(y
t
), so we need v
i
(y
tt
) > v
i
(y
t
)
so that a/b < 0. There will always be a such a y
tt
since v
i
(y
t
) < v
i
(y
t
) − . Since the
6.1. SOLVING THE M
3
N QP 81
lower-bound c = −α
i
(y
t
) < 0, again we have enough freedom to improve the objective.
Since at each iteration we are guaranteed to improve the objective if the KKT conditions
are violated and the objective is bounded, we can use the SMO in the block-coordinate
ascent algorithm to converge in a finite number of steps. To the best of our knowledge,
there are no upper bounds on the speed of convergence of SMO, but experimental evidence
has shown it a very effective algorithm for SVMs [Platt, 1999]. Of course, we can improve
the speed of convergence by adding heuristics in the selection of the pair, as long as we
guarantee that improvement is possible when KKT conditions are violated.
6.1.3 Structured SMO
Clearly, we cannot perform the above SMO updates in the space of α directly for the
structured problems, since the number of α variables is exponential. The constraints on µ
variables are much more complicated, since each µ participates not only in non-negativity
and normalization constraints, but also clique-agreement constraints. We cannot limit our
ascent steps to changing only two µ variables at a time, because in order to make a change
in one clique and stay feasible, we need to modify variables in overlapping cliques. For-
tunately, we can perform SMO updates on α variables implicitly in terms of the marginal
dual variables µ.
The diagram in Fig. 6.1.3 shows the abstract outline of the algorithm. The key steps in
the SMO algorithm are checking for violations of the KKT conditions, selecting the pair y
t
and y
tt
, computing the corresponding coefficients a, b, c, d and updating the dual. We will
show how to do these operations by doing all the hard work in terms of the polynomially
many marginal µ
i
variables and auxiliary “max-marginals” variables.
Structured KKT conditions
As before, we define v
i
(y) = w
¯
f
i
(y) +
i
(y). The KKT conditions are, for all y:
α
i
(y) = 0 ⇒ v
i
(y) ≤ v
i
(y); α
i
(y) > 0 ⇒ v
i
(y) ≥ v
i
(y). (6.5)
82 CHAPTER 6. M
3
N ALGORITHMS AND EXPERIMENTS
select
& lift
SMO
update
project
Figure 6.4: Structured SMO diagram. We use marginals µ to select an appropriate pair of
instantiations y
t
and y
tt
and reconstruct their α values. We then perform the simple SMO
update and project the result back onto the marginals.
Of course, we cannot check these explicitly. Instead, we define max-marginals for each
clique in the junction tree c ∈ T
(i)
and its values y
c
, as:
´ v
i,c
(y
c
) = max
y∼y
c
[w
¯
f
i
(y) +
i
(y)], ´ α
i,c
(y
c
) = max
y∼y
c
α
i
(y).
We also define ´ v
i,c
(y
c
) = max
y

c
,=y
c
´ v
i,c
(y
t
c
) = max
y,∼y
c
[w
¯
f
i
(y) +
i
(y)]. Note that we
do not explicitly represent α
i
(y), but we can reconstruct the maximum-entropy one from
the marginals µ
i
by using Eq. (5.13). Both ´ v
i,c
(y
c
) and ´ α
i,c
(y
c
) can be computed by using
the Viterbi algorithm (one pass propagation towards the root and one outwards from the
root [Cowell et al., 1999]). We can now express the KKT conditions in terms of the max-
marginals for each clique c ∈ T
(i)
and its values y
c
:
´ α
i,c
(y
c
) = 0 ⇒ ´ v
i,c
(y
c
) ≤ ´ v
i,c
(y
c
); ´ α
i,c
(y
c
) > 0 ⇒ ´ v
i,c
(y
c
) ≥ ´ v
i,c
(y
c
). (6.6)
Theorem 6.1.1 The KKT conditions in Eq. (6.5) and Eq. (6.6) are equivalent.
Proof:
Eq. (6.5) ⇒Eq. (6.6). Assume Eq. (6.5). Suppose, we have a violation of KKT1: for
some c, y
c
, ´ α
i,c
(y
c
) = 0, but ´ v
i,c
(y
c
) > ´ v
i,c
(y
c
). Since ´ α
i,c
(y
c
) = max
y∼y
c
α
i
(y) = 0,
6.1. SOLVING THE M
3
N QP 83
then α
i
(y) = 0, ∀y ∼ y
c
. Hence, by Eq. (6.5), v
i
(y) ≤ v
i
(y), ∀y ∼ y
c
. But ´ v
i,c
(y
c
) >
´ v
i,c
(y
c
) implies the opposite: there exists y ∼ y
c
such that v
i
(y) > ´ v
i,c
(y
c
), which also
implies v
i
(y) > v
i
(y), a contradiction.
Now suppose we have a violation of KKT2: for some i, y
c
, ´ α
i,c
(y
c
) > 0, but ´ v
i,c
(y
c
) <
´ v
i,c
(y
c
). Then v
i
(y) < v
i
(y), ∀y ∼ y
c
. But ´ α
i,c
(y
c
) > 0 implies there exists y ∼ y
c
such
that α
i
(y) > 0. For that y, by Eq. (6.5), v
i
(y) ≥ v
i
(y), a contradiction.
Eq. (6.6) ⇒ Eq. (6.5). Assume Eq. (6.6). Suppose we have a violation of KKT1:
for some y, α
i
(y) = 0, but v
i
(y) > v
i
(y). This means that y is the optimum of v
i
(),
hence ´ v
i,c
(y
c
) = v
i
(y) > v
i
(y) > ´ v
i,c
(y
c
), ∀c ∈ T
(i)
, y
c
∼ y. But by Eq. (6.6), if
´ v
i,c
(y
c
) > ´ v
i,c
(y
c
), then we cannot have ´ α
i,c
(y
c
) = 0. Hence all the y-consistent α
i
max-
marginals are positive ´ α
i,c
(y
c
) > 0, ∀c ∈ T
(i)
, and it follows that all the y-consistent
marginals µ
i
are positive as well µ
i,c
(y
c
) > 0, ∀c ∈ T
(i)
(since sum upper-bounds max).
But α
i
(y) =

c∈T
(i)
µ
i,c
(y
c
)

c∈S
(i)
µ
i,s
(y
s
)
, so if all the y-consistent marginals are positive, then α
i
(y) > 0,
a contradiction.
Now suppose we have a violation of KKT2: for some y, α
i
(y) > 0, but v
i
(y) <
v
i
(y). Since α
i
(y) > 0, we know that all the y-consistent α
i
max-marginals are positive
´ α
i,c
(y
c
) > 0, ∀c ∈ T
(i)
. By Eq. (6.6), ´ v
i,c
(y
c
) ≥ ´ v
i,c
(y
c
), ∀c ∈ T
(i)
. Note that trivially
max
y
v
i
(y
t
) = max(´ v
i,c
(y
t
c
), ´ v
i,c
(y
t
c
)) for any clique c and clique assignment y
t
c
. Since
´ v
i,c
(y
c
) ≥ ´ v
i,c
(y
c
), ∀c ∈ T
(i)
, then max
y
v
i
(y
t
) = ´ v
i,c
(y
c
), , ∀c ∈ T
(i)
. That is, ´ v
i,c
(y
c
)
is the optimal value. We will show that v
i
(y) = ´ v
i,c
(y
c
), a contradiction. To show that this,
we consider any two adjacent nodes in the tree T
(i)
, cliques a and b, with a separator s, and
show that ´ v
i,a∪b
(y
a∪b
) = ´ v
i,a
(y
a
) = ´ v
i,b
(y
b
). By chaining this equality from the root of the
tree to all the leaves, we get v
i
(y) = ´ v
i,c
(y
c
) for any c.
We need to introduce some more notation to deal with the two parts of the tree induced
by cutting the edge between a and b. Let ¦A, B¦ be a partition of the nodes T
(i)
(cliques
of (
(i)
) resulting from removing the edge between a and b such that a ∈ A and b ∈ B.
We denote the two subsets of an assignment y as y
A
and y
B
(with overlap at y
s
). The
value of an assignment v
i
(y) can be decomposed into two parts: v
i
(y) = v
i,A
(y
A
) +
v
i,B
(y
B
), where v
i,A
(y
A
) and v
i,B
(y
B
) only count the contributions of their constituent
cliques. Take any maximizer, y
(a)
∼ y
a
with v
i
(y
(a)
) = ´ v
i,a
(y
a
) ≥ ´ v
i,a
(y
a
) and any
maximizer y
(b)
∼ y
b
with v
i
(y
(b)
) = ´ v
i,b
(y
b
) ≥ ´ v
i,b
(y
b
), which by definition agree with y
84 CHAPTER 6. M
3
N ALGORITHMS AND EXPERIMENTS
on the intersection s. We decompose the two associated values into the corresponding parts:
v
i
(y
(a)
) = v
i
(y
(a)
A
)+v
i
(y
(a)
B
) and v
i
(y
(b)
) = v
i
(y
(b)
A
)+v
i
(y
(b)
B
). We create a newassignment
that combines the best of the two: y
(s)
= y
(b)
A
∪ y
(a)
B
. Note that v
i
(y
(s)
) = v
i
(y
(b)
A
) +
v
i
(y
(a)
B
) = ´ v
i,s
(y
s
), since we essentially fixed the intersection s and maximized over the
rest of the variables in A and B separately. Now ´ v
i,a
(y
a
) = ´ v
i,b
(y
b
) ≥ ´ v
i,s
(y
s
) since they
are optimal as we said above. Hence we have v
i
(y
(a)
A
) + v
i
(y
(a)
B
) = v
i
(y
(b)
A
) + v
i
(y
(b)
B
) ≥
v
i
(y
(b)
A
) + v
i
(y
(a)
B
) which implies that v
i
(y
(a)
A
) ≥ v
i
(y
(b)
A
) and v
i
(y
(b)
B
) ≥ v
i
(y
(a)
B
). Now we
create another assignment that clamps the value of both a and b: y
(a∪b)
= y
(a)
A
∪ y
(b)
B
. The
value of this assignment is optimal v
i
(y
(a∪b)
) = v
i
(y
(a)
A
) + v
i
(y
(b)
B
) = v
i
(y
(a)
) = v
i
(y
(b)
).
Structured SMO pair selection and update
As in multi-class problems, we will select the first variable in the pair, y
t
, corresponding to
a violated condition, while the second variable, y
tt
, to guarantee that solving Eq. (6.3) will
result in improving the objective. Having selected y
t
and y
tt
, the coefficients for the one-
variable QP in Eq. (6.4) are a = v
i
(y
t
)−v
i
(y
tt
), b = C[[f
i
(y
t
)−f
i
(y
tt
)[[
2
, c = −α
i
(y
t
), d =
α
i
(y
tt
). As before, we enforce approximate KKT conditions in the algorithm in Fig. 6.5.
We have two cases, corresponding to violation of KKT1 and violation of KKT2.
Case KKT1. ´ α
i,c
(y
t
c
) = 0 but ´ v
i,c
(y
t
c
) > ´ v
i,c
(y
t
c
)+. We have set y
t
= arg max
y∼y
c
v
i
(y),
so v
i
(y
t
) = ´ v
i,c
(y
t
c
) > ´ v
i,c
(y
t
c
) + > v
i
(y
t
) + and α
i
(y
t
) = 0. This is the case where
i, y
t
is a not support vector but should be. We would like to increase α
i
(y
t
), so we need
α
i
(y
tt
) > 0 to borrow from. There will always be a such a y
tt
(with y
tt
c
,= y
t
c
) since

y
α
i
(y) = 1 and α
i
(y
t
) = 0. We can find one by choosing y
c
for which ´ α
i,c
(y
c
) > 0,
which guarantees that for y
tt
c
= arg max
y∼y
c
α
i
(y), α
i
(y
tt
) > 0. Since v
i
(y
t
) ≥ v
i
(y
t
) +,
v
i
(y
t
) ≥ v
i
(y
tt
) +, so the linear coefficient in Eq. (6.4) is a = v
i
(y
t
) −v
i
(y
tt
) > . Hence
the unconstrained maximum is positive a/b > 0. Since the upper-bound d = α
i
(y
tt
) > 0,
we have enough freedom to improve the objective.
Case KKT2. ´ α
i,c
(y
t
c
) > 0 but ´ v
i,c
(y
t
c
) < ´ v
i,c
(y
t
c
)−. We have set y
t
= arg max
y∼y
c
α
i
(y),
so α
i
(y
t
) = ´ α
i,c
(y
t
c
) > 0 and v
i
(y
t
) < ´ v
i,c
(y
t
c
) < ´ v
i,c
(y
t
c
) − < v
i
(y
t
) − . This is the
case where i, y
t
is a support vector but should not be. We would like to decrease α
i
(y
t
),
so we need v
i
(y
tt
) > v
i
(y
t
) so that a/b < 0. There will always be a such a y
tt
since
6.2. EXPERIMENTS 85
1. Set violation = 0.
2. For each c ∈ T
(i)
, y
c
3. KKT1: If ´ α
i,c
(y
c
) = 0, and ´ v
i,c
(y
c
) > ´ v
i,c
(y
c
) + ,
4. Set y
t
c
= y
c
, y
t
= arg max
y∼y
c
v
i
(y) and violation = 1 and goto 7.
5. KKT2: If ´ α
i,c
(y
c
) > 0, and ´ v
i,c
(y
c
) < ´ v
i,c
(y
c
) −,
6. Set y
t
c
= y
c
, y
t
= arg max
y∼y
c
α
i
(y) and violation = 2 and goto 7.
7. If violation > 0,
8. For each y
c
,= y
t
c
,
9. If violation = 1 and ´ α
i,c
(y
c
) > 0,
10. Set y
tt
c
= arg max
y∼y
c
α
i
(y) and goto 13.
11. If violation = 2 and ´ v
i,c
(y
c
) > ´ v
i,c
(y
t
c
),
12. Set y
tt
= arg max
y∼y
c
v
i
(y) and goto 13.
13. Return y
t
and y
tt
.
Figure 6.5: Structured SMO pair selection.
v
i
(y
t
) < v
i
(y
t
) − . We can find one by choosing y
c
for which ´ v
i,c
(y
c
) > ´ v
i,c
(y
c
) − ,
which guarantees that for y
tt
c
= arg max
y∼y
c
v
i
(y), v
i
(y
tt
) > v
i
(y
t
) − , Since the lower-
bound c = −α
i
(y
t
) < 0, again we have enough freedom to improve the objective.
Having computed new values α
t
i
(y
t
) = α
i
(y
t
) + δ and α
t
i
(y
tt
) = α
i
(y
t
) − δ, we need
to project this change onto the marginal dual variables µ
i
. The only marginal affected are
the ones consistent with y
t
and/or y
tt
, and the change is very simple:
µ
t
i,c
(y
c
) = µ
i,c
(y
c
) + δ1I(y
c
∼ y
t
) −δ1I(y
c
∼ y
tt
).
6.2 Experiments
We selected a subset of ∼ 6100 handwritten words, with average length of ∼ 8 characters,
from 150 human subjects, from the data set collected by Kassel [1995]. Each word was
divided into characters, each character was rasterized into an image of 16 by 8 binary
86 CHAPTER 6. M
3
N ALGORITHMS AND EXPERIMENTS
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Log-Reg CRF mSVM M^3N
T
e
s
t

e
r
r
o
r

(
a
v
e
r
a
g
e

p
e
r
-
c
h
a
r
a
c
t
e
r
)
linear quadratic cubic
(a) (b)
Figure 6.6: (a) 3 example words from the OCR data set; (b) OCR: Average per-
character test error for logistic regression, CRFs, multiclass SVMs, and M
3
Ns, using linear,
quadratic, and cubic kernels.
pixels. (See Fig. 6.6(a).) In our framework, the image for each word corresponds to x, a
label of an individual character to ¸
j
, and a labeling for a complete word to ¸. Each label
¸
j
takes values from one of 26 classes ¦a, . . . , z¦.
The data set is divided into 10 folds of ∼ 600 training and ∼ 5500 testing examples.
The accuracy results, summarized in Fig. 6.6(b), are averages over the 10 folds. We im-
plemented a selection of state-of-the-art classification algorithms: independent label ap-
proaches, which do not consider the correlation between neighboring characters — lo-
gistic regression, multi-class SVMs as described in Eq. (2.9), and one-against-all SVMs
(whose performance was slightly lower than multi-class SVMs); and sequence approaches
— CRFs, and our proposed M
3
networks. Logistic regression and CRFs are both trained by
maximizing the conditional likelihood of the labels given the features, using a zero-mean
diagonal Gaussian prior over the parameters, with a standard deviation between 0.1 and
1. The other methods are trained by margin maximization. Our features for each label
¸
j
are the corresponding image of ith character. For the sequence approaches (CRFs and
M
3
), we used an indicator basis function to represent the correlation between ¸
j
and ¸
i+1
.
6.3. RELATED WORK 87
For margin-based methods (SVMs and M
3
), we were able to use kernels (both quadratic
and cubic were evaluated) to increase the dimensionality of the feature space. We used
the structured SMO algorithm with about 30-40 iterations through the data. Using these
high-dimensional feature spaces in CRFs is not feasible because of the enormous number
of parameters.
Fig. 6.6(b) shows two types of gains in accuracy: First, by using kernels, margin-based
methods achieve a very significant gain over the respective likelihood maximizing methods.
Second, by using sequences, we obtain another significant gain in accuracy. Interestingly,
the error rate of our method using linear features is 16% lower than that of CRFs, and
about the same as multi-class SVMs with cubic kernels. Once we use cubic kernels our
error rate is 45% lower than CRFs and about 33% lower than the best previous approach.
For comparison, the previously published results, although using a different setup (e.g., a
larger training set), are about comparable to those of multiclass SVMs.
6.3 Related work
The kernel-adatron [Friess et al., 1998] and voted-perceptron algorithms [Freund &Schapire,
1998] for large-margin classifiers have a similar online optimization scheme. Collins
[2001] have applied voted-perceptron to structured problems in natural language. Although
head-to-head comparisons have not been performed, it seems that, empirically, less passes
(about 30-40) are needed for our algorithm than in the perceptron literature.
Recently, the Exponentiated Gradient [Kivinen & Warmuth, 1997] algorithm has been
adopted to solve our structured QP for max-margin estimation [Bartlett et al., 2004]. Al-
though the EG algorithm has attractive convergence properties, it has yet to be shown to
learn faster than Structured SMO, particularly in the early iterations through the dataset.
6.4 Conclusion
In this chapter, we address the large (though polynomial) size of our quadratic program
using an effective optimization procedure inspired by SMO. In our experiments with the
OCRtask, our sequence model significantly outperforms other approaches by incorporating
88 CHAPTER 6. M
3
N ALGORITHMS AND EXPERIMENTS
high-dimensional decision boundaries of polynomial kernels over character images while
capturing correlations between consecutive characters. Overall, we believe that M
3
net-
works will significantly further the applicability of high accuracy margin-based methods to
real-world structured data. In the next two chapters, we apply this framework to important
classes of Markov networks for spatial and relational data.
Chapter 7
Associative Markov networks
In the previous chapter, we considered applications of sequence-structured Markov net-
works, which allow very efficient inference and learning. The chief computational bottle-
neck in applying Markov networks for other large-scale prediction problems is inference,
which is NP-hard in general networks suitable in a broad range of practical Markov network
structures, including grid-topology networks [Besag, 1986].
One can address the tractability issue by limiting the structure of the underlying net-
work. In some cases, such as the quad-tree model used for image segmentation [Bouman &
Shapiro, 1994], a tractable structure is determined in advance. In other cases (e.g., [Bach &
Jordan, 2001]), the network structure is learned, subject to the constraint that inference on
these networks is tractable. In many cases, however, the topology of the Markov network
does not allow tractable inference. For example, in hypertext, the network structure can
mirror the hyperlink graph, which is usually highly interconnected, leading to computa-
tionally intractable networks.
In this chapter, we show that optimal learning is feasible for an important subclass of
Markov networks — networks with attractive potentials. This subclass, called associa-
tive Markov networks (AMNs), contains networks of discrete variables with K labels and
arbitrary-size clique potentials with K parameters that favor the same labels for all vari-
ables in the clique. Such positive interactions capture the “guilt by association” pattern of
reasoning present in many domains, in which connected (“associated”) variables tend to
have the same label. AMNs are a natural fit object recognition and segmentation, webpage
89
90 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
classification, and many other applications.
In the max-margin estimation framework, the inference subtask is one of finding the
best joint (MAP) assignment to all of the variables in a Markov network. By contrast, other
learning tasks (e.g., maximizing the conditional likelihood of the target labels given the
features) require that we compute the posterior probabilities of different label assignments,
rather than just the MAP.
The MAP problem can naturally be expressed as an integer programming problem. We
use a linear program relaxation of this integer program in the min-max formulation. We
show that, for associative Markov networks of over binary variables (K = 2), this linear
program provides exact answers. To our knowledge, our method is the first to allowtraining
Markov networks of arbitrary connectivity and topology. For the non-binary case (K > 2),
the approximate linear programis not guaranteed to be optimal but we can bound its relative
error. Our empirical results suggest that the solutions of the resulting approximate max-
margin formulation work well in practice.
We present an AMN-based method for object segmentation of complex from 3D range
data. By constraining the class of Markov networks to AMNs, our models can be learned
efficiently and at run-time, scale up to tens of millions of nodes and edges. The proposed
learning formulation effectively and directly learns to exploit a large set of complex surface
and volumetric features, while balancing the spatial coherence modeled by the AMN.
7.1 Associative networks
Associative interactions arise naturally in the context of image processing, where nearby
pixels are likely to have the same label [Besag, 1986; Boykov et al., 1999b]. In this setting,
a common approach is to use a generalized Potts model [Potts, 1952], which penalizes
assignments that do not have the same label across the edge: φ
ij
(k, l) = λ
ij
, ∀k ,= l and
φ
ij
(k, k) = 1, where λ
ij
≤ 1.
For binary-valued Potts models, Greig et al. [1989] show that the MAP problem can be
formulated as a min-cut in an appropriately constructed graph. Thus, the MAP problem can
be solved exactly for this class of models in polynomial time. For L > 2, the MAP problem
7.2. LP INFERENCE 91
is NP-hard, but a procedure based on a relaxed linear program guarantees a factor 2 approx-
imation of the optimal solution [Boykov et al., 1999b; Kleinberg & Tardos, 1999]. Our
associative potentials extend the Potts model in several ways. Importantly, AMNs allow
different labels to have different attraction strength: φ
ij
(k, k) = λ
ij
(k), where λ
ij
(k) ≥ 1,
and φ
ij
(k, l) = 1, ∀k ,= l. This additional flexibility is important in many domains, as
different labels can have very diverse affinities. For example, foreground pixels tend to
have locally coherent values while background is much more varied.
In a second important extension, AMNs admit non-pairwise interactions between vari-
ables, with potentials over cliques involving m variables φ(µ
i1
, . . . , µ
im
). In this case, the
clique potentials are constrained to have the same type of structure as the edge potentials:
There are K parameters φ
c
(k, . . . , k) = λ
c
(k) ≥ 1 and the rest of the entries are set to 1.
In particular, using this additional expressive power, AMNs allow us to encode the pattern
of (soft) transitivity present in many domains. For example, consider the problem of pre-
dicting whether two proteins interact [Vazquez et al., 2003]; this probability may increase
if they both interact with another protein. This type of transitivity could be modeled by a
ternary clique that has high λ for the assignment with all interactions present.
More formally, we define associative functions and potentials as follows.
Definition 7.1.1 A function g : ¸ →IR is associative for a graph ( over K-ary variables if
it can be written as:
g(y) =

v∈1
K

k=1
g
v
(k)1I(y
v
= k) +

c∈(\1
K

k=1
g
c
(k)1I(y
c
= k, . . . , k); g
c
(k) ≥ 0, ∀c ∈ (¸1,
where 1 are the nodes and ( are the cliques of the graph (. A set of potentials φ(y) is
associative if φ(y) = e
g(y)
and g(y) is associative.
7.2 LP Inference
We can write an integer linear program for the problem of finding the maximum of an
associative function g(y), where we have a “marginal” variable µ
v
(k) for each node v ∈ 1
and each label k, which indicates whether node v has value k, and µ
c
(k) for each clique c
92 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
(containing more than one variable) and label k, which represents the event that all nodes
in the clique c have label k:
max

v∈1
K

k=1
µ
v
(k)g
v
(k) +

c∈(\1
K

k=1
µ
c
(k)g
c
(k) (7.1)
s.t. µ
c
(k) ∈ ¦0, 1¦, ∀c ∈ (, k;
K

k=1
µ
v
(k) = 1, ∀v ∈ 1;
µ
c
(k) ≤ µ
v
(k), ∀c ∈ ( ¸ 1, v ∈ c, k.
Note that we substitute the constraint µ
c
(k) =
_
v∈c
µ
v
(k) by linear inequality con-
straints µ
c
(k) ≤ µ
v
(k). This works because the coefficient g
c
(k) is non-negative and we
are maximizing the objective function. Hence at the optimum, µ
c
(k) = min
v
µ
v
(k) , which
is equivalent to µ
c
(k) =
_
v∈c
µ
v
(k), when µ
v
(k) are binary.
It can be shown that in the binary case, the linear relaxation of Eq. (7.1), (where the
constraints µ
c
(k) ∈ ¦0, 1¦ are replaced by µ
c
(k) ≥ 0), is guaranteed to produce an integer
solution when a unique solution exists.
Theorem 7.2.1 If K = 2, for any associative function g, the linear relaxation of Eq. (7.1)
has an integral optimal solution.
See Appendix A.2.1 for the proof. This result states that the MAP problem in binary AMNs
is tractable, regardless of network topology or clique size. In the non-binary case (L > 2),
these LPs can produce fractional solutions and we use a rounding procedure to get an
integral solution.
Theorem 7.2.2 If K > 2, for any associative function g, the linear relaxation of Eq. (7.1)
has a solution that is larger than the solution of the integer program by at most the number
of variables in the largest clique.
In the appendix, we also show that the approximation ratio of the rounding procedure is the
inverse of the size of the largest clique (e.g.,
1
2
for pairwise networks). Although artificial
examples with fractional solutions can be easily constructed by using symmetry, it seems
that in real data such symmetries are often broken. In fact, in all our experiments with
L > 2 on real data, we never encountered fractional solutions.
7.3. MIN-CUT INFERENCE 93
7.3 Min-cut inference
We can also use efficient min-cut algorithms to perform exact inference on the learned
models for K = 2 and approximate inference for K > 2. For simplicity, we focus on the
pairwise AMN case. We first consider the case of binary AMNs, and later show how to use
the local search algorithm developed by Boykov et al. [1999a] to perform (approximate)
inference in the general multi-class case. For pairwise, binary AMNs, the objective of the
integer program in Eq. (7.1) is:
max

v∈1

v
(1)g
v
(1) + µ
v
(2)g
v
(2)] +

uv∈c

uv
(1)g
uv
(1) + µ
uv
(2)g
uv
(2)]. (7.2)
7.3.1 Graph construction
We construct a graph in which the min-cut will correspond to the optimal MAP labeling
for the above objective. First, we recast the objective as minimization by simply reversing
the signs on the value of each θ.
min −

v∈1

v
(1)g
v
(1) + µ
v
(2)g
v
(2)] −

uv∈c

uv
(1)g
uv
(1) + µ
uv
(2)g
uv
(2)]. (7.3)
The graph will consist of a vertex for each node in the AMN, along with the 1 and 2
terminals. In the final (1
1
, 1
2
) cut, the 1
1
set will correspond to label 1, and the 1
2
set will
correspond to label 2. We will show how to deal with the node terms (those depending only
on a single variable) and the edge terms (those depending on a pair of variables), and then
how to combine the two.
Node terms
Consider a node term −µ
v
(1)g
v
(1) −µ
v
(2)g
v
(2). Such a term corresponds to the node po-
tential contribution to our objective function for node v. For each node term corresponding
to node v we add a vertex v to the min-cut graph. We then look at ∆
v
= g
v
(1) − g
v
(2),
and create an edge of weight [∆
v
[ from v to either 1 or 2, depending on the sign of ∆
v
.
The reason for that is that the final min-cut graph must consist of only positive weights. An
94 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
1 2 v
1 2 v
1
2
u v
Figure 7.1: Min-cut graph construction of node (left) and edge (right) terms.
example is presented in Fig. 7.3.1.
From Fig. 7.3.1, we see that if the AMN consisted of only node potentials, the graph
construction above would add an edge from each node to its more likely label. Thus if we
run min-cut, we would simply get a cut with cost 0, since for each introduced vertex we
have only one edge of positive weight to either 1 or 2, and we would always choose not to
cut any edges.
Edge Terms
Now consider an edge term of the form −µ
uv
(1)g
uv
(1) − µ
uv
(2)g
uv
(2). To construct a
min-cut graph for the edge term we will introduce two vertices u and v. We will connect
vertex u to 1 with an edge of weight g
uv
(1), connect v to 2 with an edge of weight g
uv
(2)
and connect u to v with an edge of weight g
uv
(1) + g
uv
(2). Fig. 7.3.1 shows an example.
Observe what happens when both nodes are on the 1
2
side of the cut: the value of the
min-cut is g
uv
(1), which must be less than g
uv
(2) or the min-cut would have placed them
both on the 1 side. When looking at edge terms in isolation, a cut that places each node
in different sets will not occur, but when we combine the graphs for node terms and edge
terms, such cuts will be possible.
We can take the individual graphs we created for node and edge terms and merge them
7.3. MIN-CUT INFERENCE 95
by adding edge weights together (and treating missing edges as edges with weight 0). It
can be shown that the resulting graph will represent the same objective (in the sense that
running min-cut on it will optimize the same objective) as the sum of the objectives of each
graph. Since our MAP-inference objective is simply a sumof node and edge terms, merging
the node and edge term graphs will result in a graph in which min-cut will correspond to
the MAP labeling.
7.3.2 Multi-class case
The graph construction above finds the best MAP labeling for the binary case, but in prac-
tice we would often like to handle multiple classes in AMNs. One of the most effective
algorithms for minimizing energy functions like ours is the α-expansion algorithm pro-
posed by Boykov et al. [1999a]. The algorithm performs a series of “expansion” moves
each of which involves optimization over two labels, and it can be shown that it converges
to within a factor of 2 of the global minimum.
Expansion Algorithm
Consider a current labeling µ and a particular label k ∈ 1, . . . , K. Another labeling µ
t
is
called an “α-expansion” move (following Boykov et al. [1999a]) from µ if µ
t
v
,= k implies
µ
t
v
= µ
v
(where µ
v
is the label of the node v in the AMN.) In other words, a k-expansion
from a current labeling allows each label to either stay the same, or change to k.
The α-expansion algorithm cycles through all labels k in either a fixed or random order,
and finds the new labeling whose objective has the lowest value. It terminates when there
is no α-expansion move for any label k that has a lower objective than the current labeling
(Fig. 7.2).
The key part of the algorithm is computing the best α-expansion labeling for a fixed
k and a fixed current labeling µ. The min-cut construction from earlier allows us to do
exactly that since an α-expansion move essentially minimizes a MAP-objective over two
labels: it either allows a node to retain its current label, or switch to the label α. In this
new binary problem we will let label 1 represent a node keeping its current label and label
2 will denote a node taking on the new label k. In order to construct the right coefficients
96 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
1. Begin with arbitrary labeling µ
2. Set success := 0
3. For each label k ∈ ¦1, . . . K¦
3.1 Compute ˆ µ = arg min −g(µ
t
) among µ
t
within one α-expansion of µ.
3.2 If E(ˆ µ) < E(µ), set µ := ˆ µ and success := 1
4. If success = 1 goto 2.
5. Return µ
Figure 7.2: α-expansion algorithm
for the new binary objective we need to consider several factors. Below, let θ
tk
i
and θ
tk,k
ij
denote the node and edge coefficients associated with the new binary objective:
◦ Node Potentials For each node i in the current labeling whose current label is not α,
we let θ
t0
i
= θ
y
i
i
, and θ
t1
i
= θ
α
i
, where y
i
denotes the current label of node i, and θ
y
i
denotes the coefficient in the multiclass AMN MAP objective. Note that we ignore
nodes with label α altogether since an α-expansion move cannot change their label.
◦ Edge Potentials For each edge (i, j) ∈ E whose nodes have labels different from α,
we add a new edge potential, with weights θ
t1
ij
= θ
α,α
ij
. If the two nodes of the edge
currently have the same label, we set θ
t0
ij
= θ
y
i
,y
j
ij
, and if the two nodes currently have
different labels we let θ
t0
ij
= 0. For each edge (i, j) ∈ E in which exactly one of the
nodes has label α in the current labeling, we add θ
α,α
ij
, to the node potential θ
t1
i
of the
node whose label is different from α.
After we have constructed the new binary MAP objective as above, we can apply the
min-cut construction from before to get the optimal labeling within one α-expansion from
the current one. Veksler [1999] shows that the α-expansion algorithm converges in O(N)
iterations where N is the number of nodes. As noted in Boykov et al. [1999a] and as we
have observed in our experiments, the algorithm terminates only after a few iterations with
most of the improvement occurring in the first 2-3 expansion moves.
7.4. MAX-MARGIN ESTIMATION 97
7.4 Max-margin estimation
The potentials of the AMN are once again log-linear combinations of basis functions. We
will need the following assumption to ensure that w
¯
f (x, y) is associative:
Assumption 7.4.1 Basis functions f are component-wise associative for ((x) for any (x, y).
Recall that this implies that for cliques larger than one, all basis functions evaluate to 0
for assignments where the values of the nodes are not equal and are non-negative for the
assignments where the values of the nodes are equal. To ensure that w
¯
f (x, y) is associa-
tive, it is useful to separate the basis functions with support only on nodes from those with
support on larger cliques.
Definition 7.4.2 Let
˙
f be the subset of basis functions f with support only on singleton
cliques:
˙
f = ¦f ∈ f : ∀ x ∈ A, y ∈ ¸, c ∈ ((((x)), [c[ > 1, f
c
(x
c
, y
c
) = 0¦.
Let
¨
f = f ¸
˙
f be the rest of the basis functions. Let ¦ ˙ w, ¨ w¦ = w be the corresponding
subsets of parameters.
It is easy to verify that any non-negative combination of associative functions is asso-
ciative, and any combination of basis functions with support only on singleton cliques is
also associative, so we have:
Lemma 7.4.3 w
¯
f (x, y) is associative for ((x) for any (x, y) whenever Assumption 7.4.1
holds and ¨ w ≥ 0.
We must make similar associative assumption on the loss function in order to guarantee
that the LP inference can handle it.
Assumption 7.4.4 The loss function (x
(i)
, y
(i)
, y) is associative for (
(i)
for all i.
In practice, this restriction is fairly mild, and the Hamming loss, which we use in general-
ization bounds and experiments, is associative.
98 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
Using the above Assumptions 7.4.1 and 7.4.4 and some algebra (see Appendix A.2.3
for derivation), we have the following max-margin QP for AMNs:
min
1
2
[[w[[
2
+ C

i,v∈1
(i)
ξ
i,v
(7.4)
s.t. w
¯
∆f
i,v
(k) −

c⊃v
m
i,c,v
(k) ≥
i,v
(k) −ξ
i,v
, ∀i, v ∈ 1
(i)
, k;
¨ w
¯

¨
f
i,c
(k) +

v∈c
m
i,c,v
(k) ≥
i,c
(k), ∀i, c ∈ (
(i)
¸ 1
(i)
, k;
m
i,c,v
(k) ≥ −¨ w
¯
¨
f
i,c
(y
(i)
c
)/[c[, ∀i, c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k;
¨ w ≥ 0;
where f
i,c
(k) = f
i,c
(k, . . . , k) and
i,c
(k) =
i,c
(k, . . . , k).
While this primal is more complex than the regular M
3
N factored primal in Eq. (5.4),
the basic structure of the first two sets of constraints remains the same: we have local
margin requirements and “credit” passed around through messages m
i,c,v
(k). The extra
constraints are due to the associativity constraints on the resulting model.
The dual of Eq. (7.4) (see derivation in Sec. A.2.3) is given by:
max

i,c∈(
(i)
, k
µ
i,c
(k)
i,c
(k) −
C
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,v∈1
(i)
, k
µ
i,v
(k)∆
˙
f
i,v
(k)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2

C
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¨ ν +

i,c∈(
(i)
, k
µ
i,c
(k)∆
¨
f
i,c
(k)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
s.t. µ
i,c
(k) ≥ 0, ∀i, ∀c ∈ (
(i)
, k;
K

k=1
µ
i,v
(k) = 1, ∀i, ∀v ∈ 1
(i)
;
µ
i,c
(k) ≤ µ
i,v
(k), ∀i, ∀c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k;
¨ ν ≥ 0.
In the dual, there are marginals µ for each node and clique, for each value k, similar
to Eq. (5.12). However, the constraints are different, and not surprisingly, are essentially
the constraints from the inference LP relaxation in Eq. (7.1).
7.5. EXPERIMENTS 99
The dual and primal solutions are related by
˙ w =

i,v∈1
(i)
, k
µ
i,v
(k)∆
˙
f
i,v
(k); ¨ w = ¨ ν +

i,c∈(
(i)
, k
µ
i,c
(k)∆
¨
f
i,c
(k).
The ¨ ν variables simply ensure that ¨ ware positive (if any component

i,c∈(
(i)
, k
µ
i,c
(k)∆
¨
f
i,c
(k)
is negative, maximizing the objective will force the corresponding component of ¨ ν to cancel
it out). Note that the objective can be written in terms of dot products of node basis func-
tions ∆
˙
f
i,v
(k)
¯

˙
f
j,¯ v
(
¯
k), so they can be kernelized. Unfortunately, the edge basis functions
cannot be kernelized because of the non-negativity constraint.
For K = 2, the LP inference is exact, so that Eq. (7.4) learns exact max-margin weights
for Markov networks of arbitrary topology. For K > 2, the linear relaxation leads to a
strengthening of the constraints on w by potentially adding constraints corresponding to
fractional assignments as in the case of untriangualated networks. Thus, the optimal choice
w, ξ for the original QP may no longer be feasible, leading to a different choice of weights.
However, as our experiments show, these weights tend to do well in practice.
7.5 Experiments
We applied associative Markov networks to the task of terrain classification. Terrain clas-
sification is very useful for autonomous mobile robots in real-world environments for path
planning, target detection, and as a pre-processing step for other perceptual tasks. The
Stanford Segbot Project
1
has provided us with a laser range maps of the Stanford campus
collected by a moving robot equipped with SICK2 laser sensors Fig. 7.5. The data consists
of around 35 million points, represented as 3D coordinates in an absolute frame of refer-
ence Fig. 7.5. Thus, the only available information is the location of points. Each reading
was a point in 3D space, represented by its (x, y, z) coordinates in an absolute frame of
reference. Thus, the only available information is the location of points, which was fairly
noisy because of localization errors.
Our task is to classify the laser range points into four classes: ground, building, tree,
1
Many thanks to Michael Montemerlo and Sebastian Thrun for sharing the data.
100 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
Figure 7.3: Segbot: roving robot equipped with SICK2 laser sensors.
and shrubbery. Since classifying ground points is trivial given their absolute z-coordinate
(height), we classify them deterministically by thresholding the z coordinate at a value
close to 0. After we do that, we are left with approximately 20 million non-ground points.
Each point is represented simply as a location in an absolute 3D coordinate system. The
features we use require pre-processing to infer properties of the local neighborhood of a
point, such as how planar the neighborhood is, or how much of the neighbors are close to
the ground. The features we use are invariant to rotation in the x-y plane, as well as the
density of the range scan, since scans tend to be sparser in regions farther from the robot.
Our first type of feature is based on the principal plane around it. For each point we
sample 100 points in a cube of radius 0.5 meters. We run PCA on these points to get the
plane of maximum variance (spanned by the first two principal components). We then par-
tition the cube into 3 3 3 bins around the point, oriented with respect to the principal
plane, and compute the percentage of points lying in the various sub-cubes. We use a num-
ber of features derived from the cube such as the percentage of points in the central column,
the outside corners, the central plane, etc. These features capture the local distribution well
and are especially useful in finding planes. Our second type of feature is based on a column
7.5. EXPERIMENTS 101
Figure 7.4: 3D laser scan range map of the Stanford Quad.
around each point. We take a cylinder of radius 0.25 meters, which extends vertically to
include all the points in a “column”. We then compute what percentage of the points lie in
various segments of this vertical column (e.g., between 2m and 2.5m). Finally, we also use
an indicator feature of whether or not a point lies within 2m of the ground. This feature is
especially useful in classifying shrubbery.
For training we select roughly 30 thousand points that represent the classes well: a
segment of a wall, a tree, some bushes. We considered three different models: SVM,
Voted-SVM and AMNs. All methods use the same set of features, augmented with a
quadratic kernel.
The first model is a multi-class SVM with a quadratic kernel over the above features.
This model (Fig. 7.5, right panel and Fig. 7.7, top panel) achieves reasonable performance
102 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
Figure 7.5: Terrain classification results showing Stanford Memorial Church obtained
with SVM, Voted-SVM and AMN models. (Color legend: buildings/red, trees/green,
shrubs/blue, ground/gray).
in many places, but fails to enforce local consistency of the classification predictions. For
example arches on buildings and other less planar regions are consistently confused for
trees, even though they are surrounded entirely by buildings.
We improved upon the SVM by smoothing its predictions using voting. For each point
we took its local neighborhood (we varied the radius to get the best possible results) and
assigned the point the label of the majority of its 100 neighbors. The Voted-SVM model
(Fig. 7.5, middle panel and Fig. 7.7, middle panel) performs slightly better than SVM: for
example, it smooths out trees and some parts of the buildings. Yet it still fails in areas like
arches of buildings where the SVM classifier has a locally consistent wrong prediction.
The final model is a pairwise AMN over laser scan points, with associative potentials
to ensure smoothness. Each point is connected to 6 of its neighbors: 3 of them are sampled
randomly from the local neighborhood in a sphere of radius 0.5m, and the other 3 are
sampled at random from the vertical cylinder column of radius 0.25m. It is important to
ensure vertical consistency since the SVMclassifier is wrong in areas that are higher off the
7.5. EXPERIMENTS 103
0 0.5 1 1.5 2 2.5
x 10
7
0
50
100
150
200
250
Problem size (nodes and edges)
R
u
n
n
i
n
g

t
i
m
e

(
s
e
c
o
n
d
s
)
Mincut inference performance
Figure 7.6: The running time (in seconds) of the min-cut-based inference algorithm for
different problem sizes. The problem size is the sum of the number of nodes and the
number of edges. Note the near linear performance of the algorithm and its efficiency even
for large models.
ground (due to the decrease in point density) or because objects tend to look different as we
vary their z-coordinate (for example, tree trunks and tree crowns look different). While we
experimented with a variety of edge features including various distances between points,
we found that even using only a constant feature performs well.
We trained the AMN model using CPLEX to solve the quadratic program; the train-
ing took about an hour on a Pentium 3 desktop. The inference over each segment was
performed using min-cut with α-expansion moves as described above. We used a pub-
licly available implementation of the min-cut algorithm, which uses bidirectional search
trees for augmenting paths (see Boykov and Kolmogorov [2004]). The implementation is
largely dominated by I/O time, with the actual min-cut taking less than two minutes even
for the largest segment. The performance is summarized in Fig. 7.6, and as we can see, it
is roughly linear in the size of the problem (number of nodes and number of edges).
We can see that the predictions of the AMN (Fig. 7.5, left panel and Fig. 7.7, bot-
tom panel) are much smoother: for example building arches and tree trunks are predicted
104 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
correctly. We also hand-labeled around 180 thousand points of the test set (Fig. 7.8) and
computed accuracies of the predictions shown in Fig. 7.9 (excluding ground, which was
classified by pre-processing). The differences are dramatic: SVM: 68%, Voted-SVM: 73%
and AMN: 93%. See more results, including a fly-through movie of the data, at
http://ai.stanford.edu/˜btaskar/3Dmap/.
7.6 Related work
Several authors have considered extensions to the Potts model. Kleinberg and Tardos
[1999] extend the multi-class Potts model to have more general edge potentials, under the
constraints that negative log of the edge potentials form a metric on the set of labels. They
also provide a solution based on a relaxed LP that has certain approximation guarantees.
More recently, Kolmogorov and Zabih [2002] showed how to optimize energy func-
tions containing binary and ternary interactions using graph cuts, as long as the parameters
satisfy a certain regularity condition. Our definition of associative potentials below also
satisfies the Kolmogorov and Zabih regularity condition for K = 2. However, the structure
of our potentials is simpler to describe and extend for the multi-class case. In fact, we can
extend our max-margin framework to estimate their more general potentials by expressing
inference as a linear program.
Our terrain classification approach is most closely related to work in vision applying
conditional random fields (CRFs) to 2D images. Kumar and Hebert [2003] train CRFs
using a pseudo-likelihood approximation to the distribution P(Y [ X) since estimating
the true conditional distribution is intractable. Unlike their work, our learning formulation
provides an exact and tractable optimization algorithm, as well as formal guarantees for
binary classification problems. Moreover, unlike their work, our approach can also handle
multi-class problems in a straightforward manner.
7.7 Conclusion
In this chapter, we provide an algorithm for max-margin training of associative Markov
networks, a subclass of Markov networks that allows only positive interactions between
7.7. CONCLUSION 105
related variables. Our approach relies on a linear programming relaxation of the MAP
problem, which is the key component in the quadratic program associated with the max-
margin formulation. We thus provide a polynomial time algorithm which approximately
solves the maximum margin estimation problem for any associative Markov network. Im-
portantly, our method is guaranteed to find the optimal (margin-maximizing) solution for all
binary-valued AMNs, regardless of the clique size or the connectivity. To our knowledge,
this algorithm is the first to provide an effective learning procedure for Markov networks
of such general structure.
Our results in the binary case rely on the fact that the LP relaxation of the MAP problem
provides exact solutions. In the non-binary case, we are not guaranteed exact solutions, but
we can prove constant-factor approximation bounds on the MAP solution returned by the
relaxed LP. It would be interesting to see whether these bounds provide us with guarantees
on the quality (e.g., the margin) of our learned model.
We present large-scale experiments with terrain segmentation and classification from
3D range data involving AMNs with tens of millions of nodes and edges. The class of
associative Markov networks appears to cover a large number of interesting applications.
We have explored only a computer vision application in this chapter, and consider another
one (hypertext classification) in the next. It would be very interesting to consider other
applications, such as extracting protein complexes from protein-protein interaction data, or
predicting links in relational data. The min-cut based inference is able to handle very large
networks, and it is an interesting challenge to apply the algorithm to even larger models
and develop efficient distributed implementations.
However, despite the prevalence of fully associative Markov networks, it is clear that
many applications call for repulsive potentials. While clearly we cannot introduce fully
general potentials into AMNs without running against the NP-hardness of the general prob-
lem, it would be interesting to see whether we can extend the class of networks we can learn
effectively.
106 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
Figure 7.7: Results from the SVM, Voted-SVM and AMN models.
7.7. CONCLUSION 107
Figure 7.8: Labeled part of the test set: ground truth (top) and SVM predictions (bottom).
108 CHAPTER 7. ASSOCIATIVE MARKOV NETWORKS
Figure 7.9: Predictions of the Voted-SVM (top) and AMN (bottom) models.
Chapter 8
Relational Markov networks
In the previous chapters, we have seen how sequential and spatial correlation between
labels can be exploited for tremendous accuracy gains. In many other supervised learning
tasks, the entities to be labeled are related with each other in very complex ways, not just
sequentially or spatially. For example, in hypertext classification, the labels of linked pages
are highly correlated. Astandard approach is to classify each entity independently, ignoring
the correlations between them. In this chapter, we present a framework that builds on
Markov networks and provides a flexible language for modeling rich interaction patterns in
structured data. We provide experimental results on a webpage classification task, showing
that accuracy can be significantly improved by modeling relational dependencies.
Many real-world data sets are innately relational: hyperlinked webpages, cross-citations
in patents and scientific papers, social networks, medical records, and more. Such data con-
sist of entities of different types, where each entity type is characterized by a different set
of attributes. Entities are related to each other via different types of links, and the link
structure is an important source of information.
Consider a collection of hypertext documents that we want to classify using some set
of labels. Most naively, we can use a bag of words model, classifying each webpage solely
using the words that appear on the page. However, hypertext has a very rich structure that
this approach loses entirely. One document has hyperlinks to others, typically indicating
that their topics are related. Each document also has internal structure, such as a partition
into sections; hyperlinks that emanate from the same section of the document are even
109
110 CHAPTER 8. RELATIONAL MARKOV NETWORKS
more likely to point to similar documents. When classifying a collection of documents,
these are important cues, that can potentially help us achieve better classification accuracy.
Therefore, rather than classifying each document separately, we want to provide a form of
collective classification, where we simultaneously decide on the class labels of all of the
entities together, and thereby can explicitly take advantage of the correlations between the
labels of related entities.
We propose the use of a joint probabilistic model for an entire collection of related enti-
ties. We introduce the framework of relational Markov networks (RMNs), which compactly
defines a Markov network over a relational data set. The graphical structure of an RMN is
based on the relational structure of the domain, and can easily model complex patterns over
related entities. For example, we can represent a pattern where two linked documents are
likely to have the same topic. We can also capture patterns that involve groups of links: for
example, consecutive links in a document tend to refer to documents with the same label.
As we show, the use of an undirected graphical model avoids the difficulties of defining
a coherent generative model for graph structures in directed models. It thereby allows us
tremendous flexibility in representing complex patterns.
8.1 Relational classification
Consider hypertext as a simple example of a relational domain. A relational domain is
defined by a schema, which describes entities, their attributes and relations between them.
In our domain, there are two entity types: Doc and Link. If a webpage is represented as a
bag of words, Doc would have a set of boolean attributes Doc.HasWord
k
indicating whether
the word k occurs on the page. It would also have the label attribute Doc.Label, indicating
the topic of the page, which takes on a set of categorical values. The Link entity type has
two attributes: Link.From and Link.To, both of which refer to Doc entities.
In general, a schema specifies of a set of entity types c = ¦E
1
, . . . , E
n
¦. Each type E is
associated with three sets of attributes: content attributes E.X(for example, Doc.HasWord
k
),
label attributes E.Y (for example, Doc.Label), and reference attributes E.R (for example,
Link.To). For simplicity, we restrict label and content attributes to take on categorical val-
ues. Reference attributes include a special unique key attribute E.K that identifies each
8.1. RELATIONAL CLASSIFICATION 111
entity. Other reference attributes E.R refer to entities of a single type E
t
= Range(E.R)
and take values in Domain(E
t
.K).
An instantiation 1 of a schema c specifies the set of entities 1(E) of each entity type
E ∈ c and the values of all attributes for all of the entities. For example, an instantiation
of the hypertext schema is a collection of webpages, specifying their labels, words they
contain and links between them. We will use 1.X, 1.Y and 1.R to denote the content,
label and reference attributes in the instantiation 1; 1.x, 1.y and 1.r to denote the values
of those attributes. The component 1.r, which we call an instantiation skeleton or instan-
tiation graph, specifies the set of entities (nodes) and their reference attributes (edges). A
hypertext instantiation graph specifies a set of webpages and links between them, but not
their words or labels. Taskar et al. [2001] suggest the use of probabilistic relational mod-
els (PRMs) for the collective classification task. PRMs [Koller & Pfeffer, 1998; Friedman
et al., 1999; Getoor et al., 2002] are a relational extension of Bayesian networks [Pearl,
1988]. A PRM specifies a probability distribution over instantiations consistent with a
given instantiation graph by specifying a Bayesian-network-like template-level probabilis-
tic model for each entity type. Given a particular instantiation graph, the PRM induces
a large Bayesian network over that instantiation that specifies a joint probability distribu-
tion over all attributes of all of the entities. This network reflects the interactions between
related instances by allowing us to represent correlations between their attributes.
In our hypertext example, a PRM might use a naive Bayes model for words, with a di-
rected edge between Doc.Label and each attribute Doc.HadWord
k
; each of these attributes
would have a conditional probability distribution P(Doc.HasWord
k
[ Doc.Label) associ-
ated with it, indicating the probability that word k appears in the document given each of
the possible topic labels. More importantly, a PRM can represent the inter-dependencies
between topics of linked documents by introducing an edge from Doc.Label to Doc.Label
of two documents if there is a link between them. Given a particular instantiation graph
containing some set of documents and links, the PRM specifies a Bayesian network over all
of the documents in the collection. We would have a probabilistic dependency from each
document’s label to the words on the document, and a dependency from each document’s
label to the labels of all of the documents to which it points. Taskar et al. show that this
approach works well for classifying scientific documents, using both the words in the title
112 CHAPTER 8. RELATIONAL MARKOV NETWORKS
and abstract and the citation-link structure.
However the application of this idea to other domains, such as webpages, is problematic
since there are many cycles in the link graph, leading to cycles in the induced “Bayesian
network”, which is therefore not a coherent probabilistic model. Getoor et al. [2001] sug-
gest an approach where we do not include direct dependencies between the labels of linked
webpages, but rather treat links themselves as random variables. Each two pages have a
“potential link”, which may or may not exist in the data. The model defines the probability
of the link existence as a function of the labels of the two endpoints. In this link exis-
tence model, labels have no incoming edges from other labels, and the cyclicity problem
disappears. This model, however, has other fundamental limitations. In particular, the re-
sulting Bayesian network has a random variable for each potential link — N
2
variables for
collections containing N pages. This quadratic blowup occurs even when the actual link
graph is very sparse. When N is large (e.g., the set of all webpages), a quadratic growth is
intractable. Even more problematic are the inherent limitations on the expressive power im-
posed by the constraint that the directed graph must represent a coherent generative model
over graph structures. The link existence model assumes that the presence of different
edges is a conditionally independent event. Representing more complex patterns involving
correlations between multiple edges is very difficult. For example, if two pages point to the
same page, it is more likely that they point to each other as well. Such interactions between
many overlapping triples of links do not fit well into the generative framework.
Furthermore, directed models such as Bayesian networks and PRMs are usually trained
to optimize the joint probability of the labels and other attributes, while the goal of clas-
sification is a discriminative model of labels given the other attributes. The advantage
of training a model only to discriminate between labels is that it does not have to trade
off between classification accuracy and modeling the joint distribution over non-label at-
tributes. In many cases, discriminatively trained models are more robust to violations of
independence assumptions and achieve higher classification accuracy than their generative
counterparts.
8.2. RELATIONAL MARKOV NETWORKS 113
Label 1
Label 2
Label 3
Figure 8.1: An unrolled Markov net over linked documents. The links follow a common
pattern: documents with the same label tend to link to each other more often.
8.2 Relational Markov networks
We now extend the framework of Markov networks to the relational setting. A relational
Markov network (RMN) specifies a conditional distribution over all of the labels of all
of the entities in an instantiation given the relational structure and the content attributes.
(We provide the definitions directly for the conditional case, as the unconditional case is a
special case where the set of content attributes is empty.) Roughly speaking, it specifies the
cliques and potentials between attributes of related entities at a template level, so a single
model provides a coherent distribution for any collection of instances from the schema.
For example, suppose that pages with the same label tend to link to each other, as
in Fig. 8.1. We can capture this correlation between labels by introducing, for each link, a
clique between the labels of the source and the target page. The potential on the clique will
have higher values for assignments that give a common label to the linked pages.
To specify what cliques should be constructed in an instantiation, we will define a no-
tion of a relational clique template. Arelational clique template specifies tuples of variables
in the instantiation by using a relational query language. For our link example, we can write
the template as a kind of SQL query:
SELECT doc1.Category, doc2.Category
114 CHAPTER 8. RELATIONAL MARKOV NETWORKS
FROM Doc doc1, Doc doc2, Link link
WHERE link.From = doc1.Key and link.To = doc2.Key
Note the three clauses that define a query: the FROM clause specifies the cross prod-
uct of entities to be filtered by the WHERE clause and the SELECT clause picks out the
attributes of interest. Our definition of clique templates contains the corresponding three
parts.
A relational clique template C = (F, W, S) consists of three components:
◦ F = ¦F
i
¦ — a set of entity variables, where an entity variable F
i
is of type E(F
i
).
◦ W(F.R) — a boolean formula using conditions of the form F
i
.R
j
= F
k
.R
l
.
◦ F.S ⊆ F.X∪ F.Y — a selected subset of content and label attributes in F.
For the clique template corresponding to the SQL query above, Fconsists of doc1, doc2
and link of types Doc, Doc and Link, respectively. W(F.R) is link.From = doc1.Key ∧
link.To = doc2.Key and F.S is doc1.Category and doc2.Category.
A clique template specifies a set of cliques in an instantiation 1:
C(1) ≡ ¦c = f .S : f ∈ 1(F) ∧ W(f .r)¦,
where f is a tuple of entities ¦f
i
¦ in which each f
i
is of type E(F
i
); 1(F) = 1(E(F
1
))
. . . 1(E(F
n
)) denotes the cross-product of entities in the instantiation; the clause W(f .r)
ensures that the entities are related to each other in specified ways; and finally, f .S selects
the appropriate attributes of the entities. Note that the clique template does not specify the
nature of the interaction between the attributes; that is determined by the clique potentials,
which will be associated with the template.
This definition of a clique template is very flexible, as the WHERE clause of a tem-
plate can be an arbitrary predicate. It allows modeling complex relational patterns on the
instantiation graphs. To continue our webpage example, consider another common pattern
in hypertext: links in a webpage tend to point to pages of the same category. This pattern
can be expressed by the following template:
SELECT doc1.Category, doc2.Category
8.2. RELATIONAL MARKOV NETWORKS 115
FROM Doc doc1, Doc doc2, Link link1, Link link2
WHERE link1.From = link2.From and link1.To = doc1.Key
and link2.To = doc2.Key and not doc1.Key = doc2.Key
Depending on the expressive power of our template definition language, we may be able
to construct very complex templates that select entire subgraph structures of an instantia-
tion. We can easily represent patterns involving three (or more) interconnected documents
without worrying about the acyclicity constraint imposed by directed models. Since the
clique templates do not explicitly depend on the identities of entities, the same template can
select subgraphs whose structure is fairly different. The RMN allows us to associate the
same clique potential parameters with all of the subgraphs satisfying the template, thereby
allowing generalization over a wide range of different structures.
A Relational Markov network (RMN) / = (C, Φ) specifies a set of clique templates
C and corresponding potentials Φ = ¦φ
C
¦
C∈C
to define a conditional distribution:
P(1.y [ 1.x, 1.r)
=
1
Z(1.x, 1.r)

C∈C

c∈C(7)
φ
C
(1.x
c
, 1.y
c
)
where Z(1.x, 1.r) is the normalizing partition function:
Z(1.x, 1.r) =

7.y

C∈C

c∈C(7)
φ
C
(1.x
c
, 1.y
t
c
)
.
Using the log-linear representation of potentials, φ
C
(V
C
) = exp¦w
¯
C
f
C
(V
C
)¦, we can
write
log P(1.y [ 1.x, 1.r) = w
¯
f (1.x, 1.y, 1.r) −log Z(1.x, 1.r)
where
f
C
(1.x, 1.y, 1.r) =

c∈C(7)
f
C
(1.x
c
, 1.y
c
)
is the sum over all appearances of the template C(1) in the instantiation, and f is the vector
116 CHAPTER 8. RELATIONAL MARKOV NETWORKS
of all f
C
.
Given a particular instantiation 1 of the schema, the RMN / produces an unrolled
Markov network over the attributes of entities in 1. The cliques in the unrolled network
are determined by the clique templates C. We have one clique for each c ∈ C(1), and
all of these cliques are associated with the same clique potential φ
C
. In our webpage
example, an RMN with the link basis function described above would define a Markov net
in which, for every link between two pages, there is an edge between the labels of these
pages. Fig. 8.1 illustrates a simple instance of this unrolled Markov network.
8.3 Approximate inference and learning
Applying both maximum likelihood and maximum margin learning in the relational setting
is requires inference in very large and complicated networks, where exact inference is
typically intractable. We therefore resort to approximate methods.
Maximum likelihood estimation
For maximum likelihood learning, we need to compute basis function expectations, not
just the most likely assignment. There is a wide variety of approximation schemes for this
problem, including MCMC and variational methods. We chose to use belief propagation
for its simplicity and relative efficiency and accuracy. Belief Propagation (BP) is a local
message passing algorithm introduced by Pearl [1988]. It is guaranteed to converge to the
correct marginal probabilities for each node only for singly connected Markov networks.
However, recent analysis [Yedidia et al., 2000] provides some theoretical justification. Em-
pirical results [Murphy et al., 1999] show that it often converges in general networks, and
when it does, the marginals are a good approximation to the correct posteriors. As our
results in Sec. 8.4 show, this approach works well in our domain. We refer the reader to
Yedidia et al. for a detailed description of the BP algorithm.
We provide a brief outline of one variant of BP, referring to [Murphy et al., 1999]
for more details. For simplicity, we assume a pairwise network where all potentials are
associated only with nodes and edges given by:
8.3. APPROXIMATE INFERENCE AND LEARNING 117
P(Y
1
, . . . , Y
n
) =
1
Z

ij
ψ
ij
(Y
i
, Y
j
)

i
ψ
i
(Y
i
)
where ij ranges over the edges of the network and ψ
ij
(Y
i
, Y
j
) = φ(x
ij
, Y
i
, Y
j
), ψ
i
(Y
i
) =
φ(x
i
, Y
i
).
The belief propagation algorithm is very simple. At each iteration, each node Y
i
sends
the following messages to all its neighbors N(i):
m
ij
(Y
j
) ←α

y
i
ψ
ij
(y
i
, Y
j

i
(y
i
)

k∈N(i)−j
m
ki
(Y
i
)
where α is a (different) normalizing constant. This process is repeated until the messages
converge. At any point in the algorithm, the marginal distribution of any node Y
i
is approx-
imated by
b
i
(Y
i
) = αψ
i
(Y
i
)

k∈N(i)
m
ki
(Y
i
)
and the marginal distribution of a pair of nodes connected by an edge is approximated by
b
ij
(Y
i
, Y
j
) = αψ
ij
(Y
i
, Y
j

i
(Y
i

j
(Y
j
)

k∈N(i)−j
m
ki
(Y
i
)

l∈N(j)−i
m
lj
(Y
j
)
These approximate marginals are precisely what we need for the computation of the
basis function expectations and performing classification. Computing the expected basis
function expectations involves summing their expected values for each clique using the
approximate marginals b
i
(Y
i
) and b
ij
(Y
i
, Y
j
). Similarly, we use max
y
i
b
i
(Y
i
) at prediction
time. Note that we can also max −product variant of loopy BP, with
m
ij
(Y
j
) ←αmax
y
i
ψ
ij
(y
i
, Y
j

i
(y
i
)

k∈N(i)−j
m
ki
(Y
i
)
to compute approximate posterior “max”-marginals and use those for prediction. In our
experiments, this results in less accurate classification, so we use posterior marginal pre-
diction.
118 CHAPTER 8. RELATIONAL MARKOV NETWORKS
Maximum margin estimation
For maximummargin estimation, we used approximate LP inference inside the max-margin
QP, using commercial Ilog CPLEX software to solve it. For networks with general poten-
tials, we used the untriangulated LP we described in Sec. 5.4. The untriangulated LP
produced fractional solutions for inference on the test data in several settings, which we
rounded independently for each label. For networks with attractive potentials (AMNs), we
used the LP in Sec. 7.2, which always produced integral solutions on test data.
8.4 Experiments
We tried out our framework on the WebKB dataset [Craven et al., 1998], which is an in-
stance of our hypertext example. The data set contains webpages from four different Com-
puter Science departments: Cornell, Texas, Washington and Wisconsin. Each page has a
label attribute, representing the type of webpage which is one of course, faculty, student,
project or other. The data set is problematic in that the category other is a grab-bag of
pages of many different types. The number of pages classified as other is quite large,
so that a baseline algorithm that simply always selected other as the label would get an
average accuracy of 75%. We could restrict attention to just the pages with the four other
labels, but in a relational classification setting, the deleted webpages might be useful in
terms of their interactions with other webpages. Hence, we compromised by eliminating
all other pages with fewer than three outlinks, making the number of other pages com-
mensurate with the other categories. The resulting category distribution is: course (237),
faculty (148), other (332), research-project (82) and student (542). The number of remain-
ing pages for each school are: Cornell (280), Texas (292), Washington (315) and Wisconsin
(454). The number of links for each school are: Cornell (574), Texas (574), Washington
(728) and Wisconsin (1614).
For each page, we have access to the entire html of the page and the links to other
pages. Our goal is to collectively classify webpages into one of these five categories. In all
of our experiments, we learn a model from three schools and test the performance of the
learned model on the remaining school, thus evaluating the generalization performance of
8.4. EXPERIMENTS 119
the different models. We used C ∈ [0.1, 10] and took the best setting for all models.
Unfortunately, we cannot directly compare our accuracy results with previous work
because different papers use different subsets of the data and different training/test splits.
However, we compare to standard text classifiers such as Naive Bayes, Logistic Regression,
and Support Vector Machines, which have been demonstrated to be successful on this data
set [Joachims, 1999].
8.4.1 Flat models
The simplest approach we tried predicts the categories based on just the text content on
the webpage. The text of the webpage is represented using a set of binary attributes that
indicate the presence of different words on the page. We found that stemming and feature
selection did not provide much benefit and simply pruned words that appeared in fewer
than three documents in each of the three schools in the training data. We also experi-
mented with incorporating meta-data: words appearing in the title of the page, in anchors
of links to the page and in the last header before a link to the page [Yang et al., 2002].
Note that meta-data, although mostly originating from pages linking into the considered
page, are easily incorporated as features, i.e. the resulting classification task is still flat
feature-based classification. Our first experimental setup compares three well-known text
classifiers — Naive Bayes, linear support vector machines (Svm), and logistic regression
(Logistic) — using words and meta-words. The results, shown in Fig. 8.2, show that the
two discriminative approaches outperformNaive Bayes. Logistic and Svmgive very sim-
ilar results. The average error over the 4 schools was reduced by around 4% by introducing
the meta-data attributes.
Incorporating meta-data gives a significant improvement, but we can take additional
advantage of the correlation in labels of related pages by classifying them collectively. We
want to capture these correlations in our model and use them for transmitting informa-
tion between linked pages to provide more accurate classification. We experimented with
several relational models.
120 CHAPTER 8. RELATIONAL MARKOV NETWORKS
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
NaïveBayes Svm Logistic
T
e
s
t

E
r
r
o
r
Words Words+Meta
Figure 8.2: Comparison of Naive Bayes, Svm, and Logistic on WebKB, with and without
meta-data features. (Only averages over the 4 schools are shown here.)
8.4.2 Link model
Our first model captures direct correlations between labels of linked pages. These corre-
lations are very common in our data: courses and research projects almost never link to
each other; faculty rarely link to each other; students have links to all categories but mostly
courses. The Link model, shown in Fig. 8.1, captures this correlation through links: in
addition to the local bag of words and meta-data attributes, we introduce a relational clique
template over the labels of two pages that are linked.We train this model using maximum
conditional likelihood (labels given the words and the links) and maximum margin.
We also compare to a directed graphical model to contrast discriminative and genera-
tive models of relational structure. The Exists-ML model is a (partially) generative model
proposed by Getoor et al. [2001]. For each page, a logistic regression model predicts
the page label given the words and meta-features. Then a simple generative model speci-
fies a probability distribution over the existence of links between pages conditioned on both
pages’ labels. Concretely, we learn the probability of existence of a link between two pages
given their labels. Note that this model does not require inference during learning. Max-
imum likelihood estimation (with regularization) of the generative component is closed
8.4. EXPERIMENTS 121
0%
5%
10%
15%
20%
25%
30%
Cor Tex Was Wis Average

E
r
r
o
r
Exists-ML
SVM
Link-ML
Link-MM
Figure 8.3: Comparison of flat versus collective classification on WebKB: SVM, Exists
model with logistic regression and the Link model estimated using the maximum likelihood
(ML) and the maximum margin (MM) criteria.
form given appropriate co-occurrence counts of linked pages’ labels. However, the predic-
tion phase is much more expensive, since the resulting graphical model includes edges not
only for the existing hyperlinks, but also those that do not exist. Intuitively, observing the
link structure directly correlates all page labels in a website, linked or not. By contrast,
the Link model avoids this problem by only modeling the conditional distribution given the
existing links.
Fig. 8.3 shows a gain in accuracy from SVMs to the Link model by using the corre-
lations between labels of linked web pages. There is also very significant additional gain
by using maximum margin training: the error rate of Link-MM is 40% lower than that of
Link-ML, and 51% lower than multi-class SVMs. The Exists model doesn’t perform very
well in comparison. This can be attributed to the simplicity of the generative model and the
difficulty of the resulting inference problem.
8.4.3 Cocite model
The second relational model uses the insight that a webpage often has internal structure
that allows it to be broken up into sections. For example, a faculty webpage might have
122 CHAPTER 8. RELATIONAL MARKOV NETWORKS
0%
5%
10%
15%
20%
25%
Cor Tex Was Wis Average

E
r
r
o
r
SVM
Cocite-ML
Cocite-MM
Figure 8.4: Comparison of Naive Bayes, Svm, and Logistic on WebKB, with and without
meta-data features. (Only averages over the 4 schools are shown here.)
one section that discusses research, with a list of links to all of the projects of the faculty
member, a second section might contain links to the courses taught by the faculty member,
and a third to his advisees. We can view a section of a webpage as a fine-grained version of
Kleinberg’s hub [Kleinberg, 1999] (a page that contains a lot of links to pages of particular
category). Intuitively, if two pages are cocited, or linked to from the same section, they are
likely to be on similar topics. Note that we expect the correlation between the labels in this
case to be positive, so we can use AMN-type potentials in the max-margin estimation. The
Cocite model captures this type of correlation.
To take advantage of this trend, we need to enrich our schema by adding the attribute
Section to Link to refer to the section number it appears in. We defined a section as a
sequence of three or more links that have the same path to the root in the html parse tree.
In the RMN, we have a relational clique template defined by:
SELECT doc1.Category, doc2.Category
FROM Doc doc1, Doc doc2, Link link1, Link link2
WHERE link1.From = link2.From and link1.Section = link2.Section and
link1.To = doc1.Key and link2.To = doc2.Key and not doc1.Key = doc2.Key
We compared the performance of SVM, Cocite-ML and Cocite-MM. The results,
8.5. RELATED WORK 123
shown in Fig. 8.4, also demonstrate significant improvements of the relational models over
the SVM. The improvement is present when testing on each of the schools. Again, maxi-
mum likelihood trained model Cocite-ML achieves a worse test error than maximum mar-
gin Cocite-MM model, which shows a 30% relative reduction in test error over SVM.
We note that, in our experiments, the learned Cocite-MM weights never produced frac-
tional solutions when used for inference, which suggests that the optimization successfully
avoided problematic parameterizations of the network, even in the case of the non-optimal
multi-class relaxation.
8.5 Related work
Our RMN representation is most closely related to the work on PRMs [Koller & Pfeffer,
1998]. Later work showed how to efficiently learn model parameters and structure (equiv-
alent of clique selection in Markov networks) from data [Friedman et al., 1999]. Getoor
et al. [2002] propose several generative models of relational structure. Their approach
easily captures the dependence of link existence on attributes of entities. However there
are many patterns that we are difficult to model in PRMs, in particular those that involve
several links at a time. We give some examples here.
One useful type of pattern type is a similarity template, where objects that share a cer-
tain graph-based property are more likely to have the same label. Consider, for example,
a professor X and two other entities Y and Z. If X’s webpage mentions Y and Z in the
same context, it is likely that the X-Y relation and the Y-Z relation are of the same type; for
example, if Y is Professor X’s advisee, then probably so is Z. Our framework accommo-
dates these patterns easily, by introducing pairwise cliques between the appropriate relation
variables.
Another useful type of subgraph template involves transitivity patterns, where the pres-
ence of an A-B link and of a B-C link increases (or decreases) the likelihood of an A-C link.
For example, students often assist in courses taught by their advisor. Note that this type
of interaction cannot be accounted for just using pairwise cliques. By introducing cliques
over triples of relations, we can capture such patterns as well. We can incorporate even
more complicated patterns, but of course we are limited by the ability of belief propagation
124 CHAPTER 8. RELATIONAL MARKOV NETWORKS
to scale up as we introduce larger cliques and tighter loops in the Markov network.
We describe and exploit these patterns in our work on RMNs using maximumlikelihood
estimation [Taskar et al., 2003b]. Attempts to model such pattern in PRMs run into the
constraint that the probabilistic dependency graph (Bayesian network) must be a directed
acyclic graph. For example, for the transitivity pattern, we might consider simply directing
the correlation edges between link existence variables arbitrarily. However, it is not clear
how to parameterize a link existence variable for a link that is involved in multiple triangles.
The structure of the relational graph has been used extensively to infer importance in
scientific publications [Egghe & Rousseau, 1990] and hypertext [Kleinberg, 1999]. Sev-
eral recent papers have proposed algorithms that use the link graph to aid classification.
Chakrabarti et al. [1998] use system-predicted labels of linked documents to iteratively
re-label each document in the test set, achieving a significant improvement compared to a
baseline of using the text in each document alone. A similar approach was used by Neville
and Jensen [2000] in a different domain. Slattery and Mitchell [2000] tried to identify di-
rectory (or hub) pages that commonly list pages of the same topic, and used these pages to
improve classification of university webpages. However, none of these approaches provide
a coherent model for the correlations between linked webpages, applying combinations of
classifiers in a procedural way, with no formal justification.
8.6 Conclusion
In this chapter, we propose a new approach for classification in relational domains. Our ap-
proach provides a coherent foundation for the process of collective classification, where we
want to classify multiple entities, exploiting the interactions between their labels. We have
shown that we can exploit a very rich set of relational patterns in classification, significantly
improving the classification accuracy over standard flat classification.
In some cases, we can incorporate relational features into standard flat classification.
For example, when classifying papers into topics, it is possible to simply view the presence
of particular citations as atomic features. However, this approach is limited in cases where
some or even all of the relational features that occur in the test data are not observed in
the training data. In our WebKB example, there is no overlap between the webpages in the
8.6. CONCLUSION 125
different schools, so we cannot learn anything from the training data about the significance
of a hyperlink to/from a particular webpage in the test data. Incorporating basic features
(e.g., words) from the related entities can aid in classification, but cannot exploit the strong
correlation between the labels of related entities that RMNs capture.
Hypertext is the most easily available source of structured data, however, RMNs are
generally applicable to any relational domain. The results in this chapter represent only
a subset of the domains we have worked on (see [Taskar et al., 2003b]). In particular,
social networks provide extensive information about interactions among people and orga-
nizations. RMNs offer a principled method for learning to predict communities of and
hierarchical structure between people and organizations based on both the local attributes
and the patterns of static and dynamic interaction. Given the wealth of possible patterns, it
is particularly interesting to explore the problem of inducing them automatically.
126 CHAPTER 8. RELATIONAL MARKOV NETWORKS
Part III
Broader applications: parsing,
matching, clustering
127
Chapter 9
Context free grammars
We present a novel discriminative approach to parsing using structured max-margin crite-
rion based on the decomposition properties of context free grammars. We show that this
framework allows high-accuracy parsing in cubic time by exploiting novel kinds of lexical
information. Our models can condition on arbitrary features of input sentences, thus incor-
porating an important kind of lexical information not usually used by conventional parsers.
We show experimental evidence of the model’s improved performance over a natural base-
line model and a lexicalized probabilistic context-free grammar.
9.1 Context free grammar model
CFGs are one of the primary formalisms for capturing the recursive structure of syntactic
constructions, although many others have also been proposed [Manning & Sch¨ utze, 1999].
For clarity of presentation, we restrict our grammars to be in Chomsky normal form as
in Sec. 3.4. The non-terminal symbols correspond to syntactic categories such as noun
phrase (NP) or verbal phrase (VP). The terminal symbols are usually words of the sen-
tence. However, in the discriminative framework that we adopt, we are not concerned with
defining a distribution over sequences of words (language model). Instead, we condition
on the words in a sentence to produce a model of the syntactic structure. Terminal sym-
bols for our purposes are part-of-speech tags like nouns (NN), verbs (VBD), determiners
(DT). For example, Fig. 9.1(a) shows a parse tree for the sentence The screen was a sea of
128
9.1. CONTEXT FREE GRAMMAR MODEL 129
red. The set of symbols we use is based on the Penn Treebank [Marcus et al., 1993]. The
non-terminal symbols with bars (for example, DT, NN, VBD) are added to conform to the
CNF restrictions. For convenience, we repeat our definition of a CFG from Sec. 3.4 here:
Definition 9.1.1 (CFG) A CFG ( consists of:
◦ A set of non-terminal symbols, ^
◦ A designated set of start symbols, ^
S
⊆ ^
◦ A set of terminal symbols, T
◦ A set of productions, T = ¦T
B
, T
U
¦, divided into
Binary productions, T
B
= ¦A →B C : A, B, C ∈ ^¦ and
Unary productions, T
U
= ¦A → D : A ∈ ^, D ∈ T ¦.
A CFG defines a set of valid parse trees in a natural manner:
Definition 9.1.2 (CFG tree) A CFG tree is a labeled directed tree, where the set of valid
labels of the internal nodes other than the root is ^ and the set of valid labels for the leaves
is T . The root’s label set is ^
S
. Additionally, each pre-leaf node has a single child and
this pair of nodes can be labeled as A and D, respectively, if and only if there is a unary
production A → D ∈ T
U
. All other internal nodes have two children, left and right, and
this triple of nodes can be labeled as A, B and C, respectively, if and only if there is a
binary production A →B C ∈ T
B
.
In general, there are exponentially many parse trees that produce a sentence of length n.
This tree representation seems quite different from the graphical models we have been
considering thus far. However, we can use an equivalent representation that essentially
encodes a tree as an assignment to a set of appropriate variables. For each span starting
with s and an ending with e, we introduce a variable Y
s,e
taking values in ^∪⊥to represent
the label of the subtree that exactly covers, or dominates, the words of the sentence from
s to e. The value ⊥ is assigned if no subtree dominates the span. Indices s and e refer to
positions between words, rather than to words themselves, hence 0 ≤ s < e ≤ n for a
sentence of length n. The “top” symbol Y
0,n
is constrained to be in ^
S
, since it represents
130 CHAPTER 9. CONTEXT FREE GRAMMARS
(a) (b)
Figure 9.1: Two representations of a binary parse tree: (a) nested tree structure, and (b)
grid of labeled spans. The row and column number are the beginning and end of the span,
respectively. Empty squares correspond to non-constituent spans. The gray squares on the
diagonal represent part-of-speech tags.
the starting symbol of the sentence. We also introduce variables Y
s,s
taking values in T to
represent the terminal symbol (part-of-speech) between s and s +1. If Y
s,e
,= ⊥, it is often
called a constituent. Fig. 9.1(b) shows the representation of the tree in Fig. 9.1(a) as a grid
where each square corresponds to Y
s,e
. The row and column number in the grid correspond
to the beginning and end of the span, respectively. Empty squares correspond to ⊥ values.
The gray squares on the diagonal represent the terminal variables. For example, the figure
shows Y
0,0
= DT, Y
6,6
= NN, Y
3,5
= NP and Y
1,4
= ⊥.
While any parse tree corresponds to an assignment to this set of variables in a straight-
forward manner, the converse is not true: there are assignments that do not correspond
to valid parse trees. In order to characterize the set of valid assignments ¸, consider the
9.1. CONTEXT FREE GRAMMAR MODEL 131
constraints that hold for a valid assignment y:
1I(y
s,e
= A) =

A→B C∈P
B
s<m<e
1I(y
s,m,e
= (A, B, C)), 0 ≤ s < e ≤ n, ∀A ∈ ^; (9.1)
1I(y
s,e
= A) =

B→A C∈P
B
0≤s

<s
1I(y
s

,s,e
= (B, A, C))
+

B→C A∈P
B
e<e

≤n
1I(y
s,e,e
= (B, C, A)), 0 ≤ s < e ≤ n, ∀A ∈ ^; (9.2)
1I(y
s,s+1
= A) =

A→D∈1
U
1I(y
s,s,s+1
= (A, D)), 0 ≤ s < n, ∀A ∈ ^; (9.3)
1I(y
s,s
= D) =

A→D∈1
U
1I(y
s,s,s+1
= (A, D)), 0 ≤ s < n, ∀D ∈ T . (9.4)
The notation y
s,m,e
= (A, B, C) abbreviates y
s,e
= A ∧ y
s,m
= B ∧ y
m,e
= C and
y
s,s,s+1
= (A, D) abbreviates y
s,s+1
= A ∧ y
s,s
= D. The first set of constraints (9.1)
holds because if the span from s to e is dominated by a subtree starting with A (that is,
y
s,e
= A), then there must be a unique production starting with A and some split point m,
s < m < e, that produces that subtree. Conversely, if y
s,e
,= A, no productions start with
A and cover s to e. The second set of constraints (9.2) holds because that if the span from
s to e is dominated by a subtree starting with A (y
s,e
= A), then there must be a (unique)
production that generated it: either starting before s or after e. Similarly, the third and
fourth set of constraints (9.3 and 9.4) hold since the terminals are generated using valid
unary productions. We denote the set of assignments y satisfying (9.1-9.4) as ¸. In fact
the converse is true as well:
Theorem 9.1.3 If y ∈ ¸, then y represents a valid CFG tree.
Proof sketch: It is straightforward to construct a parse tree from y ∈ ¸ in a top-down
manner. Starting from the root symbol, y
0,n
, the first set of constraints (9.1) ensures that a
unique production spans 0 to n, say splitting at m and specifying the values for y
0,m
and
y
m,n
. The second set of constraints (9.2) ensures that all other spans y
0,m
and y
m

,n
, for
m
t
,= m are labeled by ⊥. Recursing on the two subtrees, y
0,m
and y
m,n
, will produce the
rest of the tree down to the pre-terminals. The last two sets of constraints (9.3 and 9.4)
132 CHAPTER 9. CONTEXT FREE GRAMMARS
ensure that the terminals are generated by an appropriate unary productions from the pre-
terminals.
9.2 Context free parsing
A standard approach to parsing is to use a CFG to define a probability distribution over
parse trees. This can be done simply by assigning a probability to each production and
making sure that the sum of probabilities of all productions starting with each symbol is 1:

B,C:A→B C∈1
B
P(A →B C) = 1,

D:A→D∈1
U
P(A →D) = 1, ∀A ∈ ^.
The probability of a tree is simply the product of probabilities of the productions used in
the tree. More generally, a weighted CFG assigns a score to each production (this score
may depend on the position of the production s, m, e) such that the total score of a tree is
the sum of the score of all the productions used:
S(y) =

0≤s<m<e≤n
S
s,m,e
(y
s,m,e
) +

0≤s<n
S
s,s+1
(y
s,s+1
),
where S
s,m,e
(y
s,m,e
) = 0 if (y
s,e
= ⊥ ∨ y
s,m
= ⊥ ∨ y
m,e
= ⊥). If the production scores
are production log probabilities, then the tree score is the tree log probability. However,
weighted CFGs do not have the local normalization constraints Eq. (9.5).
We can use a Viterbi-style dynamic programming algorithm called CKY to compute
the highest score parse tree in O([T[n
3
) time [Younger, 1967; Manning & Sch¨ utze, 1999].
The algorithm computes the highest score of any subtree starting with a symbol over each
span 0 ≤ s < e ≤ n recursively:
S

s,s+1
(A) = max
A→D∈1
U
S
s,s+1
(A, D), 0 ≤ s < n, ∀A ∈ ^; (9.5)
S

s,e
(A) = max
A→B C∈P
B
s<m<e
S
s,m,e
(A, B, C) + S

s,m
(B) + S

m,e
(C), 0 ≤ s < e ≤ n, ∀A ∈ ^.
The highest scoring tree has score max
A∈A
S
S

0,n
(A). Using the arg max’s of the max’s in
9.3. DISCRIMINATIVE PARSING MODELS 133
the computation of S

, we can back-trace the highest scoring tree itself. We assume that
score ties are broken in a predetermined way, say according to some lexicographic order of
the symbols.
9.3 Discriminative parsing models
We cast parsing as a structured classification task, where we want to learn a function h :
A → ¸, where A is a set of sentences, and ¸ is a set of valid parse trees according to a
fixed CFG grammar.
The functions we consider take the following linear discriminant form:
h
w
(x) = arg max
y
w
¯
f (x, y),
where w ∈ IR
d
and f is a basis function representation of a sentence and parse tree pair
f : A ¸ → IR
d
. We assume that the basis functions decompose with the CFG structure:
f (x, y) =

0≤s≤e≤n
f (x
s,e
, y
s,e
) +

0≤s≤m<e≤n
f (x
s,m,e
, y
s,m,e
),
where n is the length of the sentence x and x
s,e
and x
s,m,e
are the relevant subsets of
the sentence the basis functions depend on. To simplify notation, we introduce the set of
indices, (, which includes both spans and span triplets:
( = ¦(s, m) : 0 ≤ s ≤ e ≤ n¦ ∪ ¦(s, m, e) : 0 ≤ s ≤ m < e ≤ n¦.
Hence, f (x, y) =

c∈(
f (x
c
, y
c
).
Note that this class of discriminants includes PCFG models, where the basis func-
tions consist of the counts of the productions used in the parse, and the parameters w are
the log-probabilities of those productions. For example, f could include functions which
identify the production used together with features of the words at positions s, m, e, and
neighboring positions in the sentence x (e.g. f(x
s,m,e
, y
s,m,e
) = 1I(y
s,m,e
= S, NP, VP) ∧
m
th
word(x) = was)). We could also include functions that identify the label of the span
134 CHAPTER 9. CONTEXT FREE GRAMMARS
from s to e together with features of the word (e.g. f(x
s,m
, y
s,m
) = 1I(y
s,m
= NP) ∧
s
th
word(x) = the)).
9.3.1 Maximum likelihood estimation
The traditional method of estimating the parameters of PCFGs assumes a generative model
that defines P(x, y) by assigning normalized probabilities to CFG productions. We then
maximize the joint log-likelihood

i
log P(x
(i)
, y
(i)
) (with some regularization). We com-
pare to such a generative grammar of Collins [1999] in our experiments.
A alternative probabilistic approach is to estimate the parameters discriminatively by
maximizing conditional log-likelihood. For example, the maximumentropy approach [John-
son, 2001] defines a conditional log-linear model:
P
w
(y [ x) =
1
Z
w
(x)
exp¦w
¯
f (x, y)¦,
where Z
w
(x) =

y
exp¦w
¯
f (x, y)¦, and maximizes the conditional log-likelihood of
the sample,

i
log P(y
(i)
[ y
(i)
), (with some regularization). The same assumption that
the basis functions decompose as sums of local functions over spans and productions is
typically made in such models. Hence, as in Markov networks, the gradient depends
on the expectations of the basis functions, which can be computed in O([T[n
3
) time by
dynamic programming algorithm called inside-outside, which is similar to the CKY al-
gorithm. However, computing the expectations over trees is actually more expensive in
practice than finding the best tree for several reasons. CKY works entirely in the log-space,
while inside-outside needs to compute actual probabilities. Branch-and-prune techniques,
which save a lot of useless computation, are only applicable in CKY.
A typical method for finding the parameters is to use Conjugate Gradients or L-BFGS
methods [Nocedal & Wright, 1999; Boyd & Vandenberghe, 2004], which repeatedly com-
pute these expectations to calculate the gradient. Clark and Curran [2004] report experi-
ments involving 479 iterations of training for one model, and 1550 iterations for another
using similar methods.
9.3. DISCRIMINATIVE PARSING MODELS 135
9.3.2 Maximum margin estimation
We assume that loss function also decomposes with the CFG structure:
(x, y, ˆ y) =

0≤s≤e≤n
(x
s,e
, y
s,e
, ˆ y
s,e
) +

0≤s≤m<e≤n
(x
s,m,e
, y
s,m,e
, ˆ y
s,m,e
) =

c∈(
(x
c
, y
c
, ˆ y
c
).
One approach would be to define (x
s,e
, y
s,e
, ˆ y
s,e
) = 1I(y
s,e
,= ˆ y
s,e
). This would lead to
(x, y, ˆ y) tracking the number of “constituent errors” in ˆ y. Another, more strict definition
would be to define (x
s,m,e
, y
s,m,e
, ˆ y
s,m,e
) = 1I(y
s,m,e
,= ˆ y
s,m,e
). This definition would lead
to (x, y, ˆ y) being the number of productions in ˆ y which are not seen in y. The constituent
loss function does not exactly correspond to the standard scoring metrics, such as F
1
or
crossing brackets, but shares the sensitivity to the number of differences between trees. We
have not thoroughly investigated the exact interplay between the various loss choices and
the various parsing metrics. We used the constituent loss in our experiments.
As in the max-margin estimation for Markov networks, we can formulate an exponen-
tial size QP:
min
1
2
[[w[[
2
+ C

i
ξ
i
(9.6)
s.t. w
¯
∆f
i
(y) ≥
i
(y) −ξ
i
∀i, y,
where ∆f
i
(y) = f (x
(i)
, y
(i)
) −f (x
(i)
, y), and
i
(y) = (x
(i)
, y
(i)
, y).
The dual of Eq. (9.6) (after normalizing by C) is given by:
max

i,y
α
i
(y)
i
(y) −
1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,y
α
i
(y)∆f
i
(y)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(9.7)
s.t.

y
α
i
(y) = 1, ∀i; α
i
(y) ≥ 0, ∀i, y.
Both of the above formulations are exponential (in the number of variables or con-
straints) in the lengths (n
i
’s) of the sentences. But we can exploit the context-free structure
of the basis functions and the loss to define a polynomial-size dual formulation in terms of
136 CHAPTER 9. CONTEXT FREE GRAMMARS
marginal variables µ
i
(y):
µ
i,s,e
(A) ≡

y:y
s,e
=A
α
i
(y), 0 ≤ s ≤ e ≤ n, ∀A ∈ ^;
µ
i,s,s
(D) ≡

y:y
s,s
=D
α
i
(y), 0 ≤ s ≤ e ≤ n, ∀D ∈ T ;
µ
i,s,m,e
(A, B, C) ≡

y:y
s,m,e
=(A,B,C)
α
i
(y), 0 ≤ s < m < e ≤ n, ∀A →B C ∈ T
B
,
µ
i,s,s,s+1
(A, D) ≡

y:y
s,s,s+1
=(A,D)
α
i
(y), 0 ≤ s < n, ∀A →D ∈ T
U
.
There are O([T
B
[n
3
i
+ [T
U
[n
i
) such variables for each sentence of length n
i
, instead of
exponentially many α
i
variables. We can now express the objective function in terms of
the marginals. Using these variables, the first set of terms in the objective becomes:

i,y
α
i
(y)
i
(y) =

i,y
α
i
(y)

c∈(
(i)

i,c
(y
c
) =

i,c∈(
(i)
,y
c
µ
i,c
(y
c
)
i,c
(y
c
).
Similarly, the second set of terms (inside the 2-norm) becomes:

i,y
α
i
(y)∆f
i
(y) =

i,y
α
i
(y)

c∈(
(i)
∆f
i,c
(y
c
) =

i,c∈(
(i)
,y
c
µ
i,c
(y
c
)∆f
i,c
(y
c
).
As in M
3
Ns, we must characterize the set of marginals µ that corresponds to valid
α. The constraints on µ are essentially based on the those that define ¸ in (9.1-9.4). In
addition, we require that the marginals over the root nodes, µ
i,0,n
i
(y
0,n
i
), sums to 1 over the
possible start symbols ^
S
.
9.3. DISCRIMINATIVE PARSING MODELS 137
Putting the pieces together, the factored dual is:
max

i,c∈(
(i)
µ
i,c
(y
c
)
i,c
(y
c
) + C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,c∈(
(i)
µ
i,c
(y
c
)∆f
i,c
(y
c
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(9.8)
s.t.

A∈A
S
µ
i,0,n
i
(A) = 1, ∀i; µ
i,c
(y
c
) ≥ 0; ∀i, ∀c ∈ (
(i)
;
µ
i,s,e
(A) =

A→B C∈P
B
s<m<e
µ
i,s,m,e
(A, B, C), ∀i, 0 ≤ s < e ≤ n
i
, ∀A ∈ ^;
µ
i,s,e
(A) =

B→A C∈P
B
0≤s

<s
µ
i,s

,s,e
(B, A, C)
+

B→C A∈P
B
e<e

≤n
i
µ
i,s,e,e
(B, C, A), ∀i, 0 ≤ s < e ≤ n
i
, ∀A ∈ ^;
µ
i,s,s+1
(A) =

A→D∈1
U
µ
i,s,s,s+1
(A, D), ∀i, 0 ≤ s < n
i
, ∀A ∈ ^;
µ
i,s,s
(D) =

A→D∈1
U
µ
i,s,s,s+1
(A, D), ∀i, 0 ≤ s < n
i
, ∀D ∈ T .
The constraints on µ is necessary, since they must correspond to marginals of a distri-
bution over trees. They are also sufficient:
Theorem 9.3.1 A set of marginals µ
i
(y) satisfying the constraints in Eq. (9.8) corresponds
to a valid distribution over the legal parse trees y ∈ ¸
(i)
. A consistent distribution α
i
(y)
is given by
α
i
(y) = µ
i,0,n
i
(y
0,n
i
)

0≤s≤m<e≤n
i
µ
i,s,m,e
(y
s,m,e
)
µ
i,s,e
(y
s,e
)
,
where 0/0 = 0 by convention.
Proof sketch: The proof follows from inside-outside probability relations [Manning &
Sch¨ utze, 1999]. The first term is a valid distribution of starting symbols. Each
µ
i,s,m,e
(y
s,m,e
)
µ
i,s,e
(y
s,e
)
term for m > s corresponds to a conditional distribution over binary productions (y
s,e

y
s,m
y
m,e
) that are guaranteed to sum to 1 over split points m and possible productions.
Similarly, each
µ
i,s,s,s+1
(y
s,s,s+1
)
µ
i,s,s+1
(y
s,s+1
)
term for corresponds to a conditional distribution over
138 CHAPTER 9. CONTEXT FREE GRAMMARS
unary productions (y
s,s+1
→ y
s,s
) that are guaranteed to sum to 1 over possible produc-
tions. Hence, we have defined a kind of PCFG (where production probabilities depend on
the location of the symbol), which induces a valid distribution α
i
over trees. It straightfor-
ward to verify that this distribution has marginals µ
i
.
9.4 Structured SMO for CFGs
We trained our max-margin models using the Structured SMOalgorithmwith block-coordinate
descent adopted from graphical models (see Sec. 6.1). The CKY algorithm computes sim-
ilar max-marginals in the course of computing the best tree as does Viterbi in Markov
networks.
´ v
i,c
(y
c
) = max
y∼y
c
[w
¯
f
i
(y) +
i
(y)], ´ α
i,c
(y
c
) = max
y∼y
c
α
i
(y).
We also define ´ v
i,c
(y
c
) = max
y

c
,=y
c
´ v
i,c
(y
t
c
) = max
y,∼y
c
[w
¯
f
i
(y) +
i
(y)]. Note that we
do not explicitly represent α
i
(y), but we can reconstruct the maximum-entropy one from
the marginals µ
i
as in Theorem 9.3.1.
We again express the KKT conditions in terms of the max-marginals for each span and
span triple c ∈ (
(i)
and its values y
c
:
´ α
i,c
(y
c
) = 0 ⇒ ´ v
i,c
(y
c
) ≤ ´ v
i,c
(y
c
); ´ α
i,c
(y
c
) > 0 ⇒ ´ v
i,c
(y
c
) ≥ ´ v
i,c
(y
c
). (9.9)
The algorithm cycles through the training sentences, runs CKY to compute the max-
marginals and performs an SMO update on the violated constraints. We typically find that
20-40 iterations through the data are sufficient for convergence in terms of the objective
function improvements.
9.5 Experiments
We used the Penn English Treebank for all of our experiments. We report results here for
each model and setting trained and tested on only the sentences of length ≤15 words. Aside
9.5. EXPERIMENTS 139
from the length restriction, we used the standard splits: sections 2-21 for training (9753
sentences), 22 for development (603 sentences), and 23 for final testing (421 sentences).
As a baseline, we trained a CNF transformation of the unlexicalized model of Klein and
Manning [2003] on this data. The resulting grammar had 3975 non-terminal symbols and
contained two kinds of productions: binary non-terminal rewrites and tag-word rewrites.
Unary rewrites were compiled into a single compound symbol, so for example a subject-
gapped sentence would have label like S+VP. These symbols were expanded back into
their source unary chain before parses were evaluated. The scores for the binary rewrites
were estimated using unsmoothed relative frequency estimators. The tagging rewrites were
estimated with a smoothed model of P(w[t), also using the model from Klein and Manning
[2003]. In particular, Table 9.2 shows the performance of this model (GENERATIVE): 87.99
F
1
on the test set.
For the BASIC max-margin model, we used exactly the same set of allowed rewrites
(and therefore the same set of candidate parses) as in the generative case, but estimated
their weights using the max-margin formulation with a loss that counts the number of
wrong spans. Tag-word production weights were fixed to be the log of the generative
P(w[t) model. That is, the only change between GENERATIVE and BASIC is the use of the
discriminative maximum-margin criterion in place of the generative maximum likelihood
one for learning production weights. This change alone results in a small improvement
(88.20 vs. 87.99 F
1
).
On top of the basic model, we first added lexical features of each span; this gave a
LEXICAL model. For a span ¸s, e) of a sentence x, the base lexical features were:
◦ x
s
, the first word in the span
◦ x
s−1
, the preceding adjacent word
◦ x
e−1
, the last word in the span
◦ x
e
, the following adjacent word
◦ ¸x
s−1
, x
s
)
◦ ¸x
e−1
, x
e
)
◦ x
s+1
for spans of length 3
140 CHAPTER 9. CONTEXT FREE GRAMMARS
Model P R F
1
GENERATIVE 87.70 88.06 87.88
BASIC 87.51 88.44 87.98
LEXICAL 88.15 88.62 88.39
LEXICAL+AUX 89.74 90.22 89.98
Table 9.1: Development set results of the various models when trained and tested on Penn
treebank sentences of length ≤ 15.
Model P R F
1
GENERATIVE 88.25 87.73 87.99
BASIC 88.08 88.31 88.20
LEXICAL 88.55 88.34 88.44
LEXICAL+AUX 89.14 89.10 89.12
COLLINS 99 89.18 88.20 88.69
Table 9.2: Test set results of the various models when trained and tested on Penn treebank
sentences of length ≤ 15.
These base features were conjoined with the span length for spans of length 3 and below,
since short spans have highly distinct behaviors (see the examples below). The features are
lexical in the sense than they allow specific words and word pairs to influence the parse
scores, but are distinct from traditional lexical features in several ways. First, there is no
notion of headword here, nor is there any modeling of word-to-word attachment. Rather,
these features pick up on lexical trends in constituent boundaries, for example the trend
that in the sentence The screen was a sea of red., the (length 2) span between the word was
and the word of is unlikely to be a constituent. These non-head lexical features capture a
potentially very different source of constraint on tree structures than head-argument pairs,
one having to do more with linear syntactic preferences than lexical selection. Regardless
of the relative merit of the two kinds of information, one clear advantage of the present
approach is that inference in the resulting model remains cubic (as opposed to O(n
5
)),
since the dynamic program need not track items with distinguished headwords. With the
addition of these features, the accuracy moved past the generative baseline, to 88.44.
9.5. EXPERIMENTS 141
As a concrete (and particularly clean) example of how these features can sway a de-
cision, consider the sentence The Egyptian president said he would visit Libya today to
resume the talks. The generative model incorrectly considers Libya today to be a base NP.
However, this analysis is counter to the trend of today to be a one-word constituent. Two
features relevant to this trend are: (CONSTITUENT ∧ first-word = today ∧ length = 1) and
(CONSTITUENT ∧ last-word = today ∧ length = 1). These features represent the preference
of the word today for being the first and last word in constituent spans of length 1.
1
In the
LEXICAL model, these features have quite large positive weights: 0.62 each. As a result,
this model makes this parse decision correctly.
Another kind of feature that can usefully be incorporated into the classification process
is the output of other, auxiliary classifiers. For this kind of feature, one must take care
that its reliability on the training not be vastly greater than its reliability on the test set.
Otherwise, its weight will be artificially (and detrimentally) high. To ensure that such
features are as noisy on the training data as the test data, we split the training into two
folds. We then trained the auxiliary classifiers on each fold, and using their predictions as
features on the other fold. The auxiliary classifiers were then retrained on the entire training
set, and their predictions used as features on the development and test sets.
We used two such auxiliary classifiers, giving a prediction feature for each span (these
classifiers predicted only the presence or absence of a bracket over that span, not bracket
labels). The first feature was the prediction of the generative baseline; this feature added
little information, but made the learning phase faster. The second feature was the output
of a flat classifier which was trained to predict whether single spans, in isolation, were
constituents or not, based on a bundle of features including the list above, but also the
following: the preceding, first, last, and following tag in the span, pairs of tags such as
preceding-first, last-following, preceding-following, first-last, and the entire tag sequence.
Tag features on the test sets were taken from a pretagging of the sentence by the tagger
described in [Toutanova et al., 2003].While the flat classifier alone was quite poor (P 78.77
/ R 63.94 / F
1
70.58), the resulting max-margin model (LEXICAL+AUX) scored 89.12 F
1
.
To situate these numbers with respect to other models, the parser in [Collins, 1999],which
1
In this length 1 case, these are the same feature. Note also that the features are conjoined with only one
generic label class “constituent” rather than specific constituent types.
142 CHAPTER 9. CONTEXT FREE GRAMMARS
is generative, lexicalized, and intricately smoothed scores 88.69 over the same train/test
configuration.
9.6 Related work
A number of recent papers have considered discriminative approaches for natural language
parsing [Johnson et al., 1999; Collins, 2000; Johnson, 2001; Geman & Johnson, 2002;
Miyao & Tsujii, 2002; Clark & Curran, 2004; Kaplan et al., 2004; Collins, 2004]. Broadly
speaking, these approaches fall into two categories, reranking and dynamic programming
approaches. In reranking methods [Johnson et al., 1999; Collins, 2000; Shen et al., 2003],
an initial parser is used to generate a number of candidate parses. A discriminative model
is then used to choose between these candidates. In dynamic programming methods, a
large number of candidate parse trees are represented compactly in a parse tree forest or
chart. Given sufficiently “local” features, the decoding and parameter estimation problems
can be solved using dynamic programming algorithms. For example, several approaches
[Johnson, 2001; Geman & Johnson, 2002; Miyao & Tsujii, 2002; Clark & Curran, 2004;
Kaplan et al., 2004] are based on conditional log-linear (maximum entropy) models, where
variants of the inside-outside algorithm can be used to efficiently calculate gradients of the
log-likelihood function, despite the exponential number of trees represented by the parse
forest.
The method we presented has several compelling advantages. Unlike reranking meth-
ods, which consider only a pre-pruned selection of “good” parses, our method is an end-
to-end discriminative model over the full space of parses. This distinction can be very
significant, as the set of n-best parses often does not contain the true parse. For example,
in the work of Collins [2000], 41% of the correct parses were not in the candidate pool of
∼30-best parses. Unlike previous dynamic programming approaches, which were based on
maximum entropy estimation, our method incorporates an articulated loss function which
penalizes larger tree discrepancies more severely than smaller ones.
Moreover, the structured SMO we use requires only the calculation of Viterbi trees,
rather than expectations over all trees (for example using the inside-outside algorithm).
9.7. CONCLUSION 143
This allows a range of optimizations that prune the space of parses (without making ap-
proximations) not possible for maximum likelihood approaches which must extract basis
function expectations from the entire set of parses. In our experiments, 20-40 iterations
were generally required for convergence (except the BASIC model, which took about 100
iterations.)
9.7 Conclusion
We have presented a maximum-margin approach to parsing, which allows a discriminative
SVM-like objective to be applied to the parsing problem. Our framework permits the use
of a rich variety of input features, while still decomposing in a way that exploits the shared
substructure of parse trees in the standard way.
It is worth considering the cost of this kind of method. At training time, discriminative
methods are inherently expensive, since they all involve iteratively checking current model
performance on the training set, which means parsing the training set (usually many times).
Generative approaches are vastly cheaper to train, since they must only collect counts from
the training set.
On the other hand, the max-margin approach does have the potential to incorporate
many new kinds of features over the input, and the current feature set allows limited lexi-
calization in cubic time, unlike other lexicalized models (including the Collins model which
it outperforms in the present limited experiments). This trade-off between the complexity,
accuracy and efficiency of a parsing model is an important area of future research.
Chapter 10
Matchings
We address the problem of learning to match: given a set of input graphs and corresponding
matchings, find a parameterized edge scoring function such that the correct matchings have
the highest score. Bipartite matchings are used in many fields, for example, to find marker
correspondences in vision problems, to map words of a sentence in one language to another,
to identify functional genetic analogues in different organisms. We have shown a compact
max-margin formulation for bipartite matchings in Ch. 4. In this chapter, we focus on a
more complex problem of non-bipartite matchings. We motivate this problem using an
application in computational biology, disulfide connectivity prediction, but non-bipartite
matchings can be used for many other tasks.
Identifying disulfide bridges formed by cysteine residues is critical in determining the
structure of proteins. Recently proposed models have formulated this prediction task as a
maximum weight perfect matching problem in a graph containing cysteines as nodes with
edge weights measuring the attraction strength of the potential bridges. We exploit combi-
natorial properties of the perfect matching problem to define a compact, convex, quadratic
program. We use kernels to efficiently learn very rich (in-fact, infinite-dimensional) mod-
els and present experiments on standard protein databases, showing that our framework
achieves state-of-the-art performance on the task.
Throughout this chapter, we use the problem of disulfide connectivity prediction as an
example. We provide some background on this problem.
144
10.1. DISULFIDE CONNECTIVITY PREDICTION 145
10.1 Disulfide connectivity prediction
Proteins containing cysteine residues form intra-chain covalent bonds known as disulfide
bridges. Such bonds are a very important feature of protein structure since they enhance
conformational stability by reducing the number of configurational states and decreasing
the entropic cost of folding a protein into its native state [Matsumura et al., 1989]. They do
so mostly by imposing strict structural constraints due to the resulting links between distant
regions of the protein sequence [Harrison & Sternberg, 1994].
Knowledge of the exact disulfide bonding pattern in a protein provides information
about protein structure and possibly its function and evolution. Furthermore, since the
disulfide connectivity pattern imposes structure constraints, it can be used to reduce the
search space in both protein folding prediction as well as protein 3D structure prediction.
Thus, the development of efficient, scalable and accurate methods for the prediction of
disulfide bonds has numerous practical applications.
Recently, there has been increased interest in applying computational techniques to
the task of predicting the intra-chain disulfide connectivity [Fariselli & Casadio, 2001;
Fariselli et al., 2002; Vullo & Frasconi, 2004; Klepeis & Floudas, 2003; Baldi et al., 2004].
Since a sequence may contain any number of cysteine residues, which may or may not
participate in disulfide bonds, the task of predicting the connectivity pattern is typically
decomposed into two subproblems: predicting the bonding state of each cysteine in the
sequence, and predicting the exact connectivity among bonded cysteines. Alternatively,
there are methods [Baldi et al., 2004] that predict the connectivity pattern without knowing
the bonding state of each cysteine
1
.
We predict the connectivity pattern by finding the maximum weighted matching in a
graph in which each vertex represents a cysteine residue, and each edge represents the
“attraction strength” between the cysteines it connects [Fariselli & Casadio, 2001]. We
parameterize the this attraction strength via a linear combination of features, which can
include the protein sequence around the two residues, evolutionary information in the form
of multiple alignment profiles, secondary structure or solvent accessibility information, etc.
1
We thank Pierre Baldi and Jianlin Cheng for introducing us to the problem of disulfide connectivity
prediction and providing us with preliminary draft of their paper and results of their model, as well as the
protein datasets.
146 CHAPTER 10. MATCHINGS
10.2 Learning to match
Formally, we seek a function h : A → ¸ that maps inputs x ∈ A to output matchings
y ∈ ¸, for example, A is the space of protein sequences and ¸ is the space of matchings
of their cysteines. The space of matchings ¸ is very large, in fact, superexponential in
the number of nodes in a graph. However, ¸ has interesting and complex combinatorial
structure which we exploit to learn h efficiently.
The training data consists of m examples S = ¦(x
(i)
, y
(i)

m
i=1
of input graphs and
output matchings. We assume that the input x defines the space of possible matchings
using some deterministic procedure. For example, given a protein sequence, we construct
a complete graph where each node corresponds to a cysteine. We represent each possible
edge between nodes j and k (j < k) in example i using a binary variable y
(i)
jk
. For simplicity,
we assume complete graphs, but very little needs to be changed to handle sparse graphs.
If example i has L
i
nodes, then there are L
i
(L
i
− 1)/2 edge variables, so y
(i)
is a
binary vector of dimension L
i
(L
i
− 1)/2. In a perfect matching, each node is connected
exactly one other node. In non-perfect matchings, each node is connected to at most one
other node. Let n
i
= L
i
/2, then for complete graphs with even number of vertices L
i
, the
number of possible perfect matchings is
(2n
i
)!
2
n
i n
i
!
(which is Ω((
n
i
2
)
n
i
), super-exponential in
n
i
). For example, 1ANS protein in Fig. 10.1 has 6 cysteines (nodes), 15 potential bonds
(edges) and 15 possible perfect matchings.
Our hypothesis class is maximum weight matchings:
h
s
(x) = arg max
y∈¸

jk
s
jk
(x)y
jk
, (10.1)
For disulfide connectivity prediction, this model was used by Fariselli and Casadio [2001].
Their model assigns an attraction strength s
jk
(x) to each pair of cysteines, calculated by
assuming that all residues in the local neighborhoods of the two cysteines make contact,
and summing contact potentials for pairs of residues. We consider a simple but very general
class of attraction scoring functions defined by a weighted combination of features or basis
functions:
s
jk
(x) =

d
w
d
f
d
(x
jk
) = w
¯
f (x
jk
), (10.2)
10.2. LEARNING TO MATCH 147
1
2 3
4
5 6
1
2 3
4
5 6
RSCCPCYWGGCPWGQNCYPEGCSGPKV
1234 56
Figure 10.1: PDB protein 1ANS: amino acid sequence, 3D structure, and graph of potential
disulfide bonds. Actual disulfide connectivity is shown in yellow in the 3D model and the
graph of potential bonds.
where x
jk
is the portion of the input x that directly relates to nodes j and k, f
d
(x
jk
) is
a real-valued basis function and w
d
∈ IR. For example, the basis functions can represent
arbitrary information about the two cysteine neighborhoods: the identity of the residues
at specific positions around the two cysteines, or the predicted secondary structure in the
neighborhood of each cysteine. We assume that the user provides the basis functions, and
that our goal is to learn the weights w, for the model:
h
w
(x) = arg max
y∈¸

jk
w
¯
f (x
jk
)y
jk
. (10.3)
Below, we will abbreviate w
¯
f (x, y) ≡

jk
w
¯
f (x
jk
)y
jk
, and w
¯
f
i
(y) ≡ w
¯
f (x
(i)
, y),
The naive formulation of the max-margin estimation, which enumerates all perfect
matchings for each example i, is:
min
1
2
[[w[[
2
s.t. w
¯
f
i
(y
(i)
) ≥ w
¯
f
i
(y) +
i
(y), ∀i, ∀y ∈ ¸
(i)
. (10.4)
The number of constraints in this formulation is super-exponential in the number of nodes
in each example. In the following sections we present two max-margin formulations,
first with an exponential set of constraints (Sec. 10.3), and then with a polynomial one
(Sec. 10.4).
148 CHAPTER 10. MATCHINGS
10.3 Min-max formulation
Using the min-max formulation from Ch. 4, we have a single max constraint for each i:
min
1
2
[[w[[
2
s.t. w
¯
f
i
(y
(i)
) ≥ max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)], ∀i. (10.5)
The key to solving this problem efficiently is the loss-augmented inference
max
y∈¸
(i) [w
¯
f
i
(y) +
i
(y)]. Under the assumption of Hamming distance loss (or any loss
function that can be written as a sum of terms corresponding to edges), this maximization
is equivalent (up to a constant term) to a maximum weighted matching problem. Note that
since the y variables are binary, the Hamming distance between y
(i)
and y can be written
as (1 − y)
¯
y
(i)
+ (1 − y
(i)
)
¯
y = 1
¯
y
(i)
+ (1 − 2y
(i)
)
¯
y. Hence, the maximum weight
matching where edge jk has weight w
¯
f (x
(i)
jk
)+(1−2y
(i)
jk
) (plus the constant 1
¯
y
(i)
) gives
the value of max
y∈¸
(i) [w
¯
f
i
(y) +
i
(y)].
This problem can be solved in O(L
3
) time [Gabow, 1973; Lawler, 1976]. It can also be
solved as a linear program, where we introduce continuous variables µ
i,jk
instead of binary
variables y
(i)
jk
.
max

jk
µ
i,jk
[w
¯
f (x
(i)
jk
) + (1 −2y
(i)
jk
)] (10.6)
s.t. µ
i,jk
≥ 0, 1 ≤ j < k ≤ L
i
;

k
µ
i,jk
≤ 1, 1 ≤ j ≤ L
i
;

j,k∈V
µ
i,jk

1
2
([V [ −1), V ⊆ ¦1, . . . , L
i
¦, [V [ ≥ 3 and odd.
The constraints

k
µ
i,jk
≤ 1 require that the number of bonds incident on a node is less
or equal to one. For perfect matchings, these constraints are changed to

k
µ
i,jk
= 1 to
ensure exactly one bond. The subset constraints (in the last line of Eq. (10.6)) ensure that
solutions to the LP are integral [Edmonds, 1965]. Note that we have an exponential number
of constraints (O(2
(L
i
−1)
)), but this number is asymptotically smaller than the number of
possible matchings . It is an open problem to derive a polynomial sized LP formulation for
perfect matchings [Schrijver, 2003].
10.3. MIN-MAX FORMULATION 149
We can write the loss-augmented inference problem in terms of the LP in Eq. (10.6):
max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)] = d
i
+ max
A
i
µ
i
≤b
i
µ
i
≥0
µ
¯
i
[F
i
w+c
i
],
where: d
i
= 1
¯
y
(i)
; µ
i
is a vector of length L
i
(L
i
− 1)/2 indexed by bond jk; A
i
and b
i
are the appropriate constraint coefficient matrix and right hand side vector, respectively. F
i
is a matrix of basis function coefficients such that the component jk of the vector F
i
w is
w
¯
f (x
(i)
jk
) and c
i
= (1 − 2y
(i)
). Note that the dependence on w is linear and occurs only
in the objective of the LP.
The dual of the LP in Eq. (10.6) is
min λ
¯
i
b
i
s.t. A
¯
i
λ
i
≥ F
i
w+c
i
; λ
i
≥ 0. (10.7)
We plug it into Eq. (10.5) and combine the minimization over λ with minimization over w.
min
1
2
[[w[[
2
(10.8)
s.t. w
¯
f
i
(y
(i)
) ≥ d
i
+ λ
¯
i
b
i
, ∀i;
A
¯
i
λ
i
≥ F
i
w+c
i
, ∀i;
λ
i
≥ 0, ∀i.
In case that our basis functions are not rich enough to predict the training data perfectly,
we can introduce a slack variable ξ
i
for each example i to allow violations of the constraints
and minimize the sum of the violations:
min
1
2
[[w[[
2
+ C

i
ξ
i
(10.9)
s.t. w
¯
f
i
(y
(i)
) + ξ
i
≥ d
i
+ λ
¯
i
b
i
, ∀i;
A
¯
i
λ
i
≥ F
i
w+c
i
, ∀i;
λ
i
≥ 0, ∀i; ξ
i
≥ 0, ∀i.
The parameter C allows the user to trade off violations of the constraints with fit to the
150 CHAPTER 10. MATCHINGS
0
5
10
15
20
25
30
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
NumberofBonds
L
o
g
-
N
u
m
b
e
r

o
f

C
o
n
s
t
r
a
i
n
t
s
PerfectMatchings
Min-max
Certificate
Figure 10.2: Log of the number of QP constraints (y-axis) vs. number of bonds (x-axis) in
the three formulations (perfect matching enumeration, min-max and certificate).
data.
Our formulation is a linearly-constrained quadratic program, albeit with an exponen-
tial number of constraints. In the next section, we develop an equivalent polynomial size
formulation.
10.4 Certificate formulation
Rather than solving the loss-augmented inference problem explicitly, we can focus on find-
ing a compact certificate of optimality that guarantees that y
(i)
= arg max
y
[w
¯
f
i
(y) +

i
(y)]. We consider perfect matchings and then provide a reduction for the non-perfect
case. Let M be a perfect matching for a complete undirected graph G = (V, E). In an
alternating cycle/path in G with respect to M, the edges alternate between those that be-
long to M and those that do not. An alternating cycle is augmenting with respect to M if
the score of the edges in the matching M is smaller that the score of the edges not in the
matching M.
Theorem 10.4.1 [Edmonds, 1965] A perfect matching M is a maximum weight perfect
matching if and only if there are no augmenting alternating cycles.
10.4. CERTIFICATE FORMULATION 151
The number of alternating cycles is exponential in the number of vertices, so simply enu-
merating all of them will not do. Instead, we can rule out such cycles by considering
shortest paths.
We begin by negating the score of those edges not in M. In the discussion below we
assume that each edge score s
jk
has been modified this way. We also refer to the score s
jk
as the length of the edge jk. An alternating cycle is augmenting if and only if its length is
negative. A condition ruling out negative length alternating cycles can be stated succinctly
using a kind of distance function. Pick an arbitrary root node r. Let d
e
j
, with j ∈ V ,
e ∈ ¦0, 1¦, denote the length of the shortest distance alternating path from r to j, where
e = 1 if the last edge of the path is in M, 0 otherwise. These shortest distances are well-
defined if and only if there are no negative alternating cycles. The following constraints
capture this distance function.
s
jk
≥ d
0
k
−d
1
j
, s
jk
≥ d
0
j
−d
1
k
, ∀ jk / ∈ M; (10.10)
s
jk
≥ d
1
k
−d
0
j
, s
jk
≥ d
1
j
−d
0
k
, ∀ jk ∈ M.
Theorem 10.4.2 There exists a distance function ¦d
e
j
¦ satisfying the constraints in
Eq. (10.10) if and only if no augmenting alternating cycles exist.
Proof.
(If) Suppose there are no augmenting alternating cycles. Since any alternating paths
from r to j can be shortened (or left the same length) by removing the cycles they contain,
the two shortest paths to j (one ending with M-edge and one not) contain no cycles. Then
let d
0
j
and d
1
j
be the length of those paths, for all j (for j = r, set d
0
r
= d
1
r
= 0). Then for
any jk (or kj) in M, the shortest path to j ending with an edge not in M plus the edge jk
(or kj) is an alternating path to k ending with an edge in M. This path is longer or same
length as the shortest path to k ending with an edge in M: s
jk
+d
0
j
≥ d
1
k
(or s
kj
+d
0
j
≥ d
1
k
),
so the constraint is satisfied. Similarly for jk, kj / ∈ M.
(Only if) Suppose a distance function ¦d
e
j
¦ satisfies the constraints in Eq. (10.10). Con-
sider an alternating cycle C. We renumber the nodes such that the cycle passes through
nodes 1, 2, . . . , l and the first edge, (1, 2), is in M. The length of the path is s(C) =
s
1,l
+

l−1
j=1
s
j,j+1
. For each odd j, the edge (j, j + 1) is in M, so s
j,j+1
≥ d
1
j+1
−d
0
j
. For
152 CHAPTER 10. MATCHINGS
even j, the edge (j, j + 1) is not in M, so s
j,j+1
≥ d
0
j+1
−d
1
j
. Finally, the last edge, (1, l),
is not in M, so s
1,l
≥ d
0
1
−d
1
l
. Summing the edges, we have:
s(C) ≥ d
0
1
−d
1
l
+
l−1

j=1,odd
[d
1
j+1
−d
0
j
] +
l−1

j=2,even
[d
0
j+1
−d
1
j
] = 0.
Hence all alternating cycles have nonnegative length.
In our learning formulation we have the loss-augmented edge weights s
(i)
jk
= (2y
(i)
jk

1)(w
¯
f (x
jk
)+1−2y
(i)
jk
). Let d
i
be a vector of distance variables d
e
j
, H
i
and G
i
be matrices
of coefficients and q
i
be a vector such that H
i
w + G
i
d
i
≥ q
i
represents the constraints
in Eq. (10.10) for example i. Then the following joint convex programin wand d computes
the max-margin parameters:
min
1
2
[[w[[
2
(10.11)
s.t. H
i
w+G
i
d
i
≥ q
i
, ∀i.
Once again, in case that our basis functions are not rich enough to predict the training data
perfectly, we can introduce a slack variable vector ξ
i
to allow violations of the constraints.
The case of non-perfect matchings can be handled by a reduction to perfect matchings
as follows [Schrijver, 2003]. We create a new graph by making a copy of the nodes and
the edges and adding edges between each node and the corresponding node in the copy.
We extend the matching by replicating its edges in the copy and for each unmatched node,
introduce an edge to its copy. We define f (x
jk
) ≡ 0 for edges between the original and
the copy. Perfect matchings in this graph projected onto the original graph correspond to
non-perfect matchings in the original graph.
The comparison between the log-number of constraints for our three equivalent QP
formulations (enumeration of all perfect matchings, min-max and certificate) is shown
in Fig. 10.2. The x-axis is the number of edges in the matching (number of nodes divided
by two).
10.5. KERNELS 153
10.5 Kernels
Instead of directly optimizing the primal problem in Eq. (10.8), we can work with its dual.
Each training example i has L
i
(L
i
− 1)/2 dual variables, and α
(i)
jk
is the dual variable
associated with the features corresponding to the edge jk. Let α
(i)
be the vector of dual
variables for example i. The dual quadratic optimization problem has the form:
max

i
c
¯
i
α
(i)

1
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i

jk∈E
(i)
__
Cy
(i)
jk
−α
(i)
jk
_
f (x
(i)
jk
)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(10.12)
s.t. A
i
α
(i)
≤ Cb
i
, ∀i.
α
(i)
≥ 0, ∀i.
The only occurrence of feature vectors is in the expansion of the squared-norm term in the
objective:

i,j

kl∈E
(i)

mn∈E
(j)
_
Cy
(i)
jk
−α
(i)
jk
_
f (x
(i)
kl
)
¯
f (x
(j)
mn
)
_
Cy
(j)
jk
−α
(j)
jk
_
(10.13)
Therefore, we can apply the kernel trick and let f (x
(i)
kl
)
¯
f (x
(j)
mn
) = K(x
(i)
kl
, x
(j)
mn
). Thus, we
can efficiently map the original features f (x
jk
) to a high-dimensional space. The primal
and dual solutions are related by:
w =

i

jk
(Cy
(i)
jk
−α
(i)
jk
)f (x
(i)
jk
) (10.14)
Eq. (10.14) can be used to compute the attraction strength s
jk
(x) in a kernelized manner at
prediction time. The polynomial-sized representation in Eq. (10.11) is similarly kerneliz-
able.
154 CHAPTER 10. MATCHINGS
10.6 Experiments
We assess the performance of our method on two datasets containing sequences with ex-
perimentally verified bonding patterns: DIPRO2 and SP39. The DIPRO2 dataset
2
was
compiled and made publicly available by Baldi et al. [2004]. It consists of all proteins
from PDB [Berman et al., 2000], as of May 2004, which contain intra-chain disulfide
bonds. After redundance reduction there are a total of 1018 sequences. In addition, the
sequences are annotated with secondary structure and solvent accessibility information de-
rived from the DSSP database [Kabsch & Sander, 1983]. The SP39 dataset is extracted
from the Swiss-Prot database of proteins [Bairoch & Apweiler, 2000], release 39. It con-
tains only sequences with experimentally verified disulfide bridges, and has a total of 726
proteins. The same dataset was used in earlier work [Baldi et al., 2004; Vullo & Frasconi,
2004; Fariselli & Casadio, 2001], and we have followed the same procedure for extracting
sequences from the database.
Even though our method is applicable to both sequences with a high number of bonds
or sequences in which the bonding state of cysteine residues is unknown, we report results
for the case where the bonding state is known, and the number of bonds is between 2 and
5 (since the case of 1 bond is trivial). The DIPRO2 contains 567 such sequences, and only
53 sequences with a higher number of bonds, so we are able to perform learning on over
90% of all proteins. There are 430 proteins with 2 and 3 bonds and 137 with 4 and 5 bonds.
SP39 contains 446 sequences containing between 2 and 5 bonds.
In order to avoid biases during testing, we adopt the same dataset splitting procedure
as the one used in previous work [Fariselli & Casadio, 2001; Vullo & Frasconi, 2004;
Baldi et al., 2004]. We split SP39 into 4 different subsets, with the constraint that pro-
teins no proteins with sequence similarity of more than 30% belong to different subsets.
Sequence similarity was derived using an all-against-all rigorous Smith-Waterman local
pairwise alignment [Smith & Waterman, 1981] (with the BLOSUM65 scoring matrix, gap
penalty 12 and gap extension 4). Pairs of chains whose alignment is less than 30 residues
were considered unrelated. The DIPRO2 dataset was split similarly into 5 folds, although
the procedure had less effect due to the redundance reduction applied by the authors of the
2
http://contact.ics.uci.edu/bridge.html
10.6. EXPERIMENTS 155
dataset.
Models
The experimental results we report use the dual formulation of Sec. 10.5 and an RBF kernel
K(x
jk
, x
lm
) = exp(
|x
jk
−x
lm
|
2
γ
), with γ ∈ [0.1, 10]. We use the exponential sized represen-
tation of Sec. 10.3 since for the case of proteins containing between two and five bonds, it
is more efficient due to the low constants in the exponential problem size. We used com-
mercial QP software (CPLEX) to train our models. Training time took around 70 minutes
for 450 examples, using a sequential optimization procedure which solves QP subproblems
associated with blocks of training examples. We are currently working on an implemen-
tation of the certificate formulation Sec. 10.4 to handle longer sequences and non-perfect
matchings (when bonding state is unknown). Below, we describe several models we used.
The features we experimented with were all based on the local regions around candidate
cysteine pairs. For each pair of candidate cysteines ¦j, k¦, where j < k, we extract the
amino-acid sequence in windows of size n centered at j and k. As in Baldi et al. [2004],
we augment the features of each model with the number of residues between j and k. The
models below use windows of size n = 9.
The first model, SEQUENCE, uses the features described above: for each window, the
actual sequence is expanded to a 20 n binary vector, in which the entries denote whether
or not a particular amino acid occurs at the particular position. For example, the 21st entry
in the vector represents whether or not the amino-acid Alanine occurs at position 2 of the
local window, counting from the left end of the window. The final set of features for each
¦j, k¦ pair of cysteines is simply the two local windows concatenated together, augmented
with the linear distance between the cysteine residues.
The second model, PROFILE, is the same as SEQUENCE, except that instead of us-
ing the actual protein sequence, we use multiple sequence alignment profile information.
Multiple alignments were computed by running PSI-BLAST using default settings to align
the sequence with all sequences in the NR database [Altschul et al., 1997]. Thus, the in-
put at each position of a local window is the frequency of occurrence of each of the 20
amino-acids in the alignments.
156 CHAPTER 10. MATCHINGS
K PROFILE DAG-RNN
2 0.75 / 0.75 0.74 / 0.74
3 0.60 / 0.48 0.61 / 0.51
4 0.46 / 0.24 0.44 / 0.27
5 0.43 / 0.16 0.41 / 0.11
K SEQUENCE PROFILE PROFILE-SS
2 0.70 / 0.70 0.73 / 0.73 0.79 / 0.79
3 0.62 / 0.52 0.67 / 0.59 0.74 / 0.69
4 0.44 / 0.21 0.59 / 0.44 0.70 / 0.56
5 0.29 / 0.06 0.43 / 0.17 0.62 / 0.27
(a) (b)
Table 10.1: Numbers indicate Precision / Accuracy. (a) Performance of PROFILE model
on SP39 vs. preliminary results of the DAG-RNN model [Baldi et al., 2004] which repre-
sent the best currently published results. In each row, the best performance is in bold. (b)
Performance of SEQUENCE, PROFILE, PROFILE-SS models on the DIPRO2 dataset.
The third model, PROFILE-SS, augments the PROFILE model with secondary structure
and solvent-accessibility information. The DSSP program produces 8 types of secondary
structure, so we augment each local window of size n with an additional length 8 n
binary vector, as well as a length n binary vector representing the solvent accessibility at
each position.
Results and discussion
We evaluate our algorithm using two metrics: accuracy and precision. The accuracy mea-
sure counts how many full connectivity patterns were predicted correctly, whereas preci-
sion measures the number of correctly predicted bonds as a fraction of the total number of
possible bonds.
The first set of experiments compares our model to preliminary results reported in Baldi
et al. [2004], which represent the current top-performing system. We perform 4-fold cross-
validation on SP39 in order to replicate their setup. As Table 10.1 shows, the PROFILE
model achieves comparable results, with similar or better levels of precision for all bond
numbers, and slightly lower accuracies for the case of 2 and 3 bonds.
In another experiment, we show the performance gained by using multiple alignment
information by comparing the results of the SEQUENCE model with the PROFILE. As we
can see from Table 10.1(b), the evolutionary information captured by the amino-acid align-
ment frequencies plays an important role in increasing the performance of the algorithm.
10.7. RELATED WORK 157
2 and 3 bonds
20%
30%
40%
50%
60%
70%
80%
90%
100%
100 150 200 250 300 350 400 450
Training Set Size
Accuracy (2 Bonds)
Precision (3 Bonds)
Accuracy (3 Bonds)
4 and 5 bonds
5%
15%
25%
35%
45%
55%
65%
75%
85%
100 150 200 250 300 350 400 450
Training Set Size
Precision (4 Bonds)
Accuracy (4 Bonds)
Precision (5 Bonds)
Accuracy (5 Bonds)
(a) (b)
Figure 10.3: Performance of PROFILE model as training set size changes for proteins with
(a) 2 and 3 bonds (b) 4 and 5 bonds.
The same phenomenon is observed by Vullo and Frasconi [2004] in their comparison of
sequence and profile-based models.
As a final experiment, we examine the role that secondary structure and solvent-accessibility
information plays in the model PROFILE-SS. Table 10.1(b) shows that the gains are sig-
nificant, especially for sequences with 3 and 4 bonds. This highlights the importance of
developing even richer features, perhaps through more complex kernels.
Fig. 10.3 shows the performance of the PROFILE model as training set size grows. We
can see that for sequences of all bond numbers, both accuracy and precision increase as the
amount of data grows. The trend is more pronounced for sequences with 4 and 5 bonds
because they are sparsely distributed in the dataset. Such behavior is very promising, since
it validates the applicability of our algorithm as the availability of high-quality disulfide
bridge annotations increases with time.
10.7 Related work
The problem of inverse perfect matching has been studied by Liu and Zhang [2003] in
the inverse combinatorial optimization framework we describe in Sec. 4.3: Given a set of
nominal weights w
0
and a perfect matching M, which is not a maximum one with respect
to w
0
, find a new weight vector w that makes M optimal and minimizes [[w
0
− w[[
p
for p = 1, ∞. They do not provide a compact optimization problem for this related but
158 CHAPTER 10. MATCHINGS
different task, relying instead on the ellipsoid method with constraint generation.
The problem of disulfide bond prediction first received comprehensive computational
treatment in Fariselli and Casadio [2001]. They modeled the prediction problem as finding
a perfect matching in a weighted graph where vertices represent bonded cysteine residues,
and edge weights correspond to attraction strength. The problem of learning the edge
weights was addressed using a simulated annealing procedure. Their method is only ap-
plicable to the case when bonding state is known. In Fariselli et al. [2002], the authors
switch to using a neural network for learning edge weights and achieve better performance,
especially for the case of 2 and 3 disulfide bonds.
The method in Vullo and Frasconi [2004] takes a different approach to the problem. It
scores candidate connectivity patterns according to their similarity with respect to the cor-
rect pattern, and uses a recursive neural network architecture [Frasconi et al., 1998] to score
candidate patterns. At prediction time the pattern scores are used to perform an exhaustive
search on the space of all matchings. The method is computationally limited to sequences
of 2 to 5 bonds. It also uses multiple alignment profile information and demonstrates its
benefits over sequence information.
In Baldi et al. [2004], the authors achieve the current state-of-the-art performance on
the task. Their method uses Directed Acyclic Graph Recursive Neural Networks [Baldi
& Pollastri, 2003] to predict bonding probabilities between cysteine pairs. The prediction
problem is solved using a weighted graph matching based on these probabilities. Their
method performs better than the one in Vullo and Frasconi [2004] and is also the only
one which can cope with sequences with more than 5 bonds. It also improves on previous
methods by not assuming knowledge of bonding state.
A different approach to predicting disulfide bridges is reported in Klepeis and Floudas
[2003], where bond prediction occurs as part of predicting β-sheet topology in proteins.
Residue-to-residue contacts (which include disulphide bridges) are predicted by solving a
series of constrained integer programming problems. Interestingly, the approach can be
used to predict disulfide bonds with no knowledge of bonding state, but the results are not
comparable with those in other publications.
The task of predicting whether or not a cysteine is bonded has also been addressed using
a variety of machine learning techniques including neural networks, SVMs, and HMMs
10.8. CONCLUSION 159
[Fariselli et al., 1999; Fiser & Simon, 2000; Martelli et al., 2002; Frasconi et al., 2002;
Ceroni et al., 2003] Currently the top performing systems have accuracies around 85%.
10.8 Conclusion
In this chapter, we derive a compact convex quadratic program for the problem of learning
to match. Our approach learns a parameterized scoring function that reproduces the ob-
served matchings in a training set. We present two formulations: one which is based on a
linear programming approach to matching, requiring an exponential number of constraints,
and one which develops a certificate of matching optimality for a compact polynomial-sized
representation. We apply our framework to the task of disulfide connectivity prediction, for-
mulated as a weighted matching problem. Our experimental results show that the method
can achieve performance comparable to current top-performing systems. Furthermore, the
use of kernels makes it easy to incorporate rich sets of features such as secondary structure
information, or extended local neighborhoods of the protein sequence. In the future, it will
be worthwhile to examine how other kernels, such as convolution kernels for protein se-
quences, will affect performance. We also hope to explore the more challenging problem
of disulfide connectivity prediction when the bonding state of cysteines is unknown. While
we have developed the framework to handle that task, it remains to experimentally deter-
mine how well the method performs, especially in comparison to existing methods [Baldi
et al., 2004], which have already addressed the more challenging setting.
Chapter 11
Correlation clustering
Data can often be grouped in many different reasonable clusterings. For example, one user
may organize her email messages by project and time, another by sender and topic. Images
can be segmented by hue or object boundaries. For a given application, there might be
only one of these clusterings that is desirable. Learning to cluster considers the problem of
finding desirable clusterings on new data, given example desirable clusterings on training
data.
We focus on correlation clustering, a novel clustering method that has recently en-
joyed significant attention from the theoretical computer science community [Bansal et al.,
2002; Demaine & Immorlica, 2003; Emanuel & Fiat, 2003]. It is formulated as a vertex
partitioning problem: Given a graph with real-valued edge scores (both positive and neg-
ative), partition the vertices into clusters to maximize the score of intra-cluster edges, and
minimize the weight of inter-cluster edges. Positive edge weights represent correlations be-
tween vertices, encouraging those vertices to belong to a common cluster; negative weights
encourage the vertices to belong to different clusters. Unlike most clustering formulations,
correlation clustering does not require the user to specify the number of clusters nor a dis-
tance threshold for clustering; both of these parameters are effectively chosen to be the best
possible by the problem definition. These properties make correlation clustering a promis-
ing approach to many clustering problems; in machine learning, it has been successfully
applied to coreference resolution for proper nouns [McCallum & Wellner, 2003].
Recently, several algorithms based on linear programming and positive-semidefinite
160
11.1. CLUSTERING FORMULATION 161
programming relaxations have been proposed to approximately solve this problem. In this
chapter, we employ these relaxations to derive a max-margin formulation for learning the
edge scores for correlation clustering from clustered training data. We formulate the ap-
proximate learning problem as a compact convex program with quadratic objective and
linear or positive-semidefinite constraints. Experiments on synthetic and real-world data
show the ability of the algorithm to learn an appropriate clustering metric for a variety of
desired clusterings.
11.1 Clustering formulation
An instance of correlation clustering is specified by an undirected graph ( = (1, c) with
N nodes and edge score s
jk
for each jk in c, (j < k). We assume that the graph is fully
connected (if it is not, we can make it fully connected by adding appropriate edges jk with
s
jk
= 0). We define binary variables y
jk
, one for each edge jk, that represent whether node
j and k belong to the same cluster. Let ¸ be the space of assignments y that define legal
partitions. For notational convenience, we introduce both y
jk
and y
kj
variables, which will
be constrained to have the same value. We also introduce y
jj
variables, and fix them to
have value 1 and set s
jj
= 0.
Bansal et al. [2002] consider two related problems:
max
y∈¸

jk:s
jk
>0
s
jk
y
jk

jk:s
jk
<0
s
jk
(1 −y
jk
); (MAXAGREE)
min
y∈¸

jk:s
jk
>0
s
jk
(1 −y
jk
) −

jk:s
jk
<0
s
jk
y
jk
; (MINDISAGREE)
The motivation for the names of the two problems comes from separating the set of edges
into positive weight edges and negative weight edges. The best score is obviously achieved
by including all the positive and excluding all the negative edges, but this will not generally
produce a valid partition. In MAXAGREE, we maximize the “agreement” of the partition
with the positive/negative designations: the weight of the positive included edges minus
the weight of negative excluded edges. In MINDISAGREE, we minimize the disagreement:
the weight of positive excluded edges minus the weight of negative included edges. In
162 CHAPTER 11. CORRELATION CLUSTERING
particular, let
s

= max
y∈¸

s
jk
y
jk
; s

=

jk:s
jk
<0
s
jk
; s
+
=

jk:s
jk
>0
s
jk
.
Then the value of MAXAGREE is s

− s

and the value of MINDISAGREE s
+
− s

. The
optimal partition for the two problems is of course the same (if it is unique). Bansal
et al. [2002] show that both of these problems are NP-hard (but have different approxi-
mation hardness). We will concentrate on the maximization version, MAXAGREE. Several
approximation algorithms have been developed based on Linear and Semidefinite Program-
ming [Charikar et al., 2003; Demaine & Immorlica, 2003; Emanuel & Fiat, 2003], which
we consider in the next sections.
11.1.1 Linear programming relaxation
In order to insure that y defines a partition, it is sufficient to enforce a kind of triangle
inequality for each triple of nodes j < k < l:
y
jk
+ y
kl
≤ y
jl
+ 1; y
jk
+ y
jl
≤ y
kl
+ 1; y
jl
+ y
kl
≤ y
jk
+ 1. (11.1)
The triangle inequality enforces transitivity: if j and k are in the same cluster (y
jk
= 1)
and k and l are in the same cluster (y
kl
= 1), then j and l will be forced to be in this
cluster (y
jl
= 1). The other two cases are similar. Any symmetric, transitive binary relation
induces a partition of the objects.
With these definitions, we can express the MAXAGREE problem as an integer linear
program (ignoring the constant −s

):
max

jk
s
jk
y
jk
(11.2)
s.t. y
jk
+ y
kl
≤ y
lj
+ 1, ∀j, k, l; y
jj
= 1, ∀j; y
jk
∈ ¦0, 1¦, ∀j, k.
Note that the constraints imply that y
ab
= y
ba
for any two nodes a and b. To see this,
consider the inequalities involving node a and b with j = a, k = b, l = b and j = b, k =
11.1. CLUSTERING FORMULATION 163
a, l = a:
y
ab
+ y
bb
≤ y
ba
+ 1; y
ba
+ y
aa
≤ y
ab
+ 1;
Since y
aa
= y
bb
= 1, we have y
ab
= y
ba
.
The LP relaxation is obtained by replacing the binary variables y
jk
∈ ¦0, 1¦ in Eq. (11.2)
with continuous variables 0 ≤ µ
jk
≤ 1.
µ
jk
+ µ
kl
≤ µ
lj
+ 1, ∀j, k, l; µ
jj
= 1, ∀j; µ
jk
≥ 0, ∀j, k. (11.3)
Note that µ
jk
≤ 1 is implied since the triangle inequality with j = l gives µ
jk
+ µ
kj

µ
jj
+ 1, and since µ
jj
= 1 and µ
jk
= µ
kj
, we have µ
jk
≤ 1.
We are not guaranteed that this relaxation will produce integral solutions. The LP
solution, s
LP
, is an upper bound on the s

. Charikar et al. [2003] show that the integrality
gap of this upper bound is at least 2/3:
s

−s

s
LP
−s

<
2
3
.
11.1.2 Semidefinite programming relaxation
An alternative formulation [Charikar et al., 2003] is the SDP relaxation. Let mat(µ) denote
the variables µ
jk
arranged into a matrix.
max

jk
s
jk
µ
jk
(11.4)
s.t. mat(µ) _ 0; µ
jj
= 1, ∀j; µ
jk
≥ 0, ∀j, k.
In effect, we substituted the triangle inequalities by the semidefinite constraint. To motivate
this relaxation, consider any clustering solution. Choose a collection of orthogonal unit
vectors ¦v
1
, . . . , v
K
¦, one for each cluster in the solution. Every vertex j in the cluster
is assigned the unit vector v
j
corresponding to the cluster it is in. If vertices j and k
are in the same cluster, then v
¯
j
v
k
= 1, if not, v
¯
j
v
k
= 0. The score of the clustering
164 CHAPTER 11. CORRELATION CLUSTERING
solution can now be expressed in terms of the dot products v
¯
j
v
k
. In the SDP relaxation
in Eq. (11.4), we have mat(µ) _ 0, which can be decomposed into a sum of outer products
mat(µ) =

j
v
j
v
¯
j
.
The entries µ
jk
correspond to inner products v
¯
j
v
k
. The vectors generating the inner
products are unit vectors by the requirement µ
jj
= 1. However, they do not necessarily
form a set of orthogonal vectors.
The SDP solution, s
SDP
, is also an upper bound on the s

and Charikar et al. [2003]
show that the integrality gap of this upper bound is at least 0.82843:
s

−s

s
SDP
−s

≤ 0.82843.
11.2 Learning formulation
The score for an edge is commonly derived from the characteristics of the pair of nodes.
Specifically, we parameterize the score as a weighted combination of basis functions
s
jk
= w
¯
f (x
jk
),
w, f (x
jk
) ∈ IR
n
, where x
jk
is a set of features associated with nodes j and k. In document
clustering, the entries of f
x
jk
might be the words shared by the documents j and k, while if
one is clustering points in IR
n
, features might be distances along different dimensions. We
assume f (x
jj
) = 0 so that s
jj
= 0. Hence we write
w
¯
f (x, y) =

jk
y
jk
w
¯
f (x
jk
).
Furthermore, we assume that the loss function decomposes over the edges, into a sum
of edge losses
i,jk
(y
jk
):

i
(y) =

jk

i,jk
(y
jk
) =

jk
y
jk

i,jk
(1) + (1 −y
jk
)
i,jk
(0) =
i
(0) +

jk
y
jk

i,jk
,
where
i,jk
=
i,jk
(1) −
i,jk
(0). For example, the Hamming loss counts the number of
11.2. LEARNING FORMULATION 165
edges incorrectly cut or uncut by a partition.

H
i
(y) =

jk
1I(y
jk
,= y
(i)
jk
) =

jk
y
(i)
jk
+

jk
y
jk
(1 −2y
(i)
jk
) =
H
i
(0) +

jk
y
jk

i,jk
,
where
i,jk
= 1 −2y
(i)
jk
.
With this assumption, the loss augmented maximization is

i
(0) + max
y∈¸

jk
y
jk
[w
¯
f (x
(i)
jk
) +
i,jk
]. (11.5)
We can now use the LP relaxation in Eq. (11.3) and the SDP relaxation in Eq. (11.4) as
upper bounds on Eq. (11.5). We use these upper-bounds in the min-max formulation to
achieve approximate max-margin estimation.
The dual of the LP based upper bound for example i is
i
(0)+
min

jkl
λ
i,jkl
+

j
z
i,j
(11.6)
s.t.

l

i,jkl
+ λ
i,ljk
−λ
i,klj
] ≥ w
¯
f (x
(i)
jk
) +
i,jk
, ∀j ,= k;
z
i,j
+

l

i,jjl
+ λ
i,ljj
−λ
i,jlj
] ≥
i,jj
, ∀j;
λ
i,jkl
≥ 0, ∀j, k, l.
Above, we introduced a dual variable λ
i,jkl
for each triangle inequality and z
i,j
for the
identity on the diagonal. Note the righthand-side of the second set of inequalities follows
from the assumption f (x
jj
) = 0.
166 CHAPTER 11. CORRELATION CLUSTERING
Plugging this dual into the min-max formulation, we have:
min
1
2
[[w[[
2
+ C

ξ
i
(11.7)
s.t. w
¯
f (x
(i)
, y
(i)
) + ξ
i

i
(0) +

jkl
λ
i,jkl
+

j
z
i,j
, ∀i;

l

i,jkl
+ λ
i,ljk
−λ
i,klj
] ≥ w
¯
f (x
(i)
jk
) +
i,jk
, ∀i, ∀j ,= k;
z
i,j
+

l

i,jjl
+ λ
i,ljj
−λ
i,jlj
] ≥
i,jj
, ∀i, ∀j;
λ
i,jkl
≥ 0, ∀i, ∀j, k, l.
Similarly, the dual of the SDP based upper bound is
i
(0)+
min

j
z
i,j
(11.8)
s.t. −λ
i,jk
≥ w
¯
f (x
(i)
jk
) +
i,jk
, ∀j ,= k;
z
i,j
−λ
i,jj

i,jj
∀j;
mat(λ
i
) _ 0.
Plugging the SDP dual into the min-max formulation, we have:
min
1
2
[[w[[
2
+ C

ξ
i
(11.9)
s.t. w
¯
f (x
(i)
, y
(i)
) + ξ
i

i
(0) +

j
z
i,j
, ∀i;
−λ
i,jk
≥ w
¯
f (x
(i)
jk
) +
i,jk
, ∀i, ∀j ,= k;
z
i,j
−λ
i,jj

i,jj
, ∀i, ∀j;
mat(λ
i
) _ 0, ∀i, ∀j, k, l.
11.3. DUAL FORMULATION AND KERNELS 167
11.3 Dual formulation and kernels
The dual of Eq. (11.7) and Eq. (11.9) provide some insight into the structure of the problem
and enable efficient use of kernels. Here we give the dual of Eq. (11.7):
max

i,jk
µ
i,jk

i,jk

1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,jk
(y
(i)
jk
−µ
i,jk
)f (x
(i)
jk
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
s.t. µ
jk
+ µ
kl
≤ µ
lj
+ 1, ∀i, ∀j, k, l; µ
i,jj
= 1, ∀i, ∀j; µ
i,jk
≥ 0, ∀i, ∀j, k.
The dual of Eq. (11.9) is very similar, except that the linear transitivity constraints
µ
jk
+ µ
kl
≤ µ
lj
+ 1, ∀i, ∀j, k, l are replaced by the corresponding mat(µ
i
) _ 0:
max

i,jk
µ
i,jk

i,jk

1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,jk
(y
(i)
jk
−µ
i,jk
)f (x
(i)
jk
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
s.t. mat(µ
i
) _ 0, ∀i; µ
i,jj
= 1, ∀i, ∀j; µ
i,jk
≥ 0, ∀i, ∀j, k.
The relation between the primal and dual solution is
w = C

i,jk
(y
(i)
jk
−µ
jk
)f (x
(i)
jk
). (11.10)
One important consequence of this relationship is that the edge parameters are all sup-
port vector expansions. The dual objective can be expressed in terms of dot-products
f (x
jk
)
¯
f (x
lm
). Therefore, we can use kernels K(x
jk
, x
lm
) to define the space of basis
functions. This kernel looks at two pairs of nodes, (j, k) and (l, m), and measures the
similarity between the relation between the nodes of each pair. If we are clustering points
in Euclidian space, the kernel could be a function of the two segments corresponding to
the pairs of points, for example, a polynomial kernel over their lengths and angle between
them.
168 CHAPTER 11. CORRELATION CLUSTERING
11.4 Experiments
We present experiments on a synthetic problem exploring the effect of irrelevant basis
functions (features), and two real data sets, email clustering and image segmentation.
11.4.1 Irrelevant features
In this section, we explore on a synthetic example how our algorithm deals with irrele-
vant features. In particular, we generate data (100 points) from a mixture of two one-
dimensional Gaussians, where each mixture component corresponds to a cluster. This first
dimension is thus the relevant feature. Then we add noise components in D additional
(irrelevant) dimensions. The noise is independently generated for each dimension, from a
mixture of Gaussians with same difference in means and variance as for the relevant dimen-
sion. Figure 11.1(a) shows the projection of a data sample onto the first two dimensions
and on two irrelevant dimensions.
Let x
j
denote each point and x
j
[d] denote the d-th dimension of the point. We used
a basis function for each dimension f
d
(x
jk
) = e
−(x
j
[d]−x
k
[d])
2
, plus an additional constant
basis function. The training and test data consists of a 100 samples from the model. The
results in Fig. 11.1(b) illustrate the capability of our algorithm to learn to ignore irrelevant
dimensions. The accuracy is the fraction of edges correctly predicted to be between/within
cluster. Random clustering will give an accuracy of 50%. The comparison with k-means is
simply a baseline to illustrate the effect of the noise on the data.
11.4.2 Email clustering
We also test our approach on the task of clustering email into folders. We gathered the
data from the SRI CALO project.
1
Our dataset consisted of email from seven users (ap-
proximately 100 consecutive messages per user), which the users had manually filed into
different folders. The number of folders for each users varied from two to six, with an
average of 3-4 folders. We are interested in the problem of learning to cluster emails.
1
http://www.ai.sri.com/project/CALO
11.4. EXPERIMENTS 169
−4 −3 −2 −1 0 1 2 3
−4
−3
−2
−1
0
1
2
3
4
−4 −3 −2 −1 0 1 2 3 4
−4
−3
−2
−1
0
1
2
3
4
0.45
0.55
0.65
0.75
0.85
0.95
1 2 4 8 16 32 64 128
#noisedimensions
t
e
s
t

a
c
c
u
r
a
c
y
lcc
k-means
(a) (b)
Figure 11.1: (a) Projection onto first two dimensions (top) and two noise dimensions (bot-
tom); (c) Performance on 2 cluster problem as function of the number of irrelevant noise
dimensions. Learning to cluster is the solid line, k-means the dashed line. Error-bars de-
note one standard deviation, averages are over 20 runs. Accuracy is the fraction of edges
correctly predicted to be between/within cluster.
Specifically, what score function s
jk
causes correlation clustering to give clusters similar
to those that human users had chosen?
To test our learning algorithm, we use each user as a training set in turn, learning the
parameters from the partition of a single user’s mailbox into folders. We then use the
learned parameters to cluster the other users’ mail. The basis functions f (x
jk
) measured the
similarity between the text of the messages, the similarity between the “From:” field, “To:”
field, and “Cc:” field. One feature was used for each common word in the pair of emails
(except words that appeared in more than half the messages, which were deemed “stop
words” and omitted). Also, additional features captured the proportion of shared tokens
for each email field, including the from, to, Cc, subject and body fields. The algorithm
is therefore able to automatically learn the relative importance of certain email fields to
filing two messages together, as well as importance of meaningful words versus common,
irrelevant ones.
We compare our method to the k-means clustering algorithm using with the same word
170 CHAPTER 11. CORRELATION CLUSTERING
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
1 2 3 4 5 6 7
User
P
a
i
r

F
1
lcc
k-means
Figure 11.2: Average Pair F1 measure for clustering user mailboxes.
features, and took the best clustering out of five tries. We made this comparison somewhat
easy for k-means by giving it the correct number of clusters k. We also informed our algo-
rithm of the number of clusters by uniformly adding a positive weight to the edge weights
to cause it to give the correct number of clusters. We performed a simple binary search on
this additional bias weight parameter to find the number of clusters comparable to k. The
results in Fig. 11.2 show the average F1 measure (harmonic mean of precision and recall;
a standard metric in information retrieval [Baeza-Yates & Ribeiro-Neto, 1999]) computed
on the pairs of messages that belonged to the same cluster. Our algorithm significantly
outperforms k-means on several users and does worse only for one of the users.
11.4.3 Image segmentation
We also test our approach on the task of image segmentation. We selected images from
the Berkeley Image Segmentation Dataset [Martin et al., 2001] for which two users had
significantly different segmentations. For example, Fig. 11.3(a) and (b) show two distinct
segmentations: one very coarse, mostly based of overall hue, and one much finer, based
on the hue and intensity. Depending on the task at hand, we may prefer the first over the
second or vice-versa. It is precisely this kind of variability in the similarity judgements that
11.4. EXPERIMENTS 171
(a) (b)
Figure 11.3: Two segmentations by different users: training image with (a) coarse segmen-
tation and (b) fine segmentation.
we want our algorithm to capture.
In order to segment the image, we first divided it contiguous regions of approximately
the same color by running connected components on the pixels. We connected two adjacent
pixels by an edge if their RGB value was the same at a coarse resolution (4 bits per each
of the R,G,B channel). We then selected about a hundred largest regions, which covered
80-90% of the pixels. These regions are the objects that our algorithm learns to cluster.
(We then use the learned metric to greedily assign the remaining small regions to the large
adjoining regions.)
There is a rich space of possible features we can use in our models: for each pair
of regions, we can consider their shape, color distribution, distance, presence of edges
between them, etc. In our experiments, we used a fairly simple set of features that are very
easy to compute. For each region, we calculated the bounding box, area and average color
(averaging the pixels in the RGB space). We then computed three distances (one for each
HSV channel), as well as the distance in pixels between the bounding boxes and the area
of the smaller of the two regions. All features were normalized to have zero mean and
variance 1.
We trained two models, one using Fig. 11.3(a) and the other using Fig. 11.3(b) and
172 CHAPTER 11. CORRELATION CLUSTERING
(a) (b) (c)
Figure 11.4: Test image: (a) input; (b) segmentation based on coarse training data (c)
segmentation based on fine training data.
tested on the image in Fig. 11.4(a). The results are shown in Fig. 11.4(b) and (c), respec-
tively. Note that mountains, rocks and grass are segmented very coarsely based on hue in
(b) while the segmentation in (c) is more detailed and sensitive to saturation and value of
the colors.
11.5 Related work
The performance of most clustering algorithms depends critically on the distance metric
that they are given for measuring the similarity or dissimilarity between different data-
points. Recently, a number of algorithms have been proposed for automatically learning
distance metrics as a preprocessing step for clustering [Xing et al., 2002; Bar-Hillel et al.,
2003]. In contrast to algorithms that learn a metric independently of the algorithm that will
be used to cluster the data, we describe a formulation that tightly integrates metric learning
with the clustering algorithm, tuning one to the other in a joint optimization. Thus, instead
of using an externally-defined criterion for choosing the metric, we will instead seek to
learn a good metric for the clustering algorithm. An example of work in a similar vein is
the algorithm for learning a distance metric for spectral clustering [Bach & Jordan, 2003].
The clustering algorithm essentially uses an eigenvector decomposition of an appropriate
matrix derived from the pairwise affinity matrix, which is more efficient than correlation
clustering, for which we use LP or SDP formulations. However the objective in the learning
11.6. CONCLUSION 173
formulation proposed in Bach and Jordan [2003] is not convex and difficult to optimize.
11.6 Conclusion
We looked at correlation clustering, and how to learn the edge weights from example clus-
terings. Our approach ties together the inference and learning algorithm, and attempts
to learn a good metric specifically for the clustering algorithm.. We showed results on a
synthetic dataset, showcasing robustness to noise dimensions. Experiments on the CALO
e-mail and image segmentation experiments show the potential of the algorithm on real-
world data. The main limitation of the correlation clustering is scalability: the number of
constraints ([1[
3
) in the LP relaxation and the size of the positive-semidefinite constraint
in the SDP relaxation. It would be very interesting to explore constraint generation or sim-
ilar approaches to speed up learning and inference. On the theoretical side, it would be
interesting to work out a PAC-like bound for generalization of the learned score metric.
174 CHAPTER 11. CORRELATION CLUSTERING
Part IV
Conclusions and future directions
175
Chapter 12
Conclusions and future directions
This thesis presents a novel statistical estimation framework for structured models based
on the large margin principle underlying support vector machines. The framework results
in several efficient learning formulations for complex prediction tasks. Fundamentally,
we rely on the expressive power of convex optimization problems to compactly capture
inference or solution optimality in structured models. Directly embedding this structure
within the learning formulation produces compact convex problems for efficient estimation
of very complex models. For some of these models, alternative estimation methods are
intractable. We develop theoretical foundations for our approach and show a wide range
of experimental applications, including handwriting recognition, 3D terrain classification,
disulfide connectivity in protein structure prediction, hypertext categorization, natural lan-
guage parsing, email organization and image segmentation.
12.1 Summary of contributions
We view a structured prediction model as a mapping from the space of inputs x ∈ A to
a discrete vector output y ∈ ¸. Essentially, a model defines a compact, parameterized
scoring function w
¯
f (x, y) and prediction using the model reduces to finding the highest
scoring output y given the input x. Our class of models has the following linear form:
h
w
(x) = arg max
y: g(x,y)≤0
w
¯
f (x, y),
176
12.1. SUMMARY OF CONTRIBUTIONS 177
where w ∈ IR
n
is the vector of parameters of the model, constraints g(x, y) ∈ IR
k
define
the space of feasible outputs y given the input x and basis functions f (x, y) ∈ IR
n
represent
salient features of the input/output pair. Although the space of outputs ¦y : g(x, y) ≤ 0¦ is
usually immense, we assume that the inference problem arg max
y : g(x,y)≤0
w
¯
f (x, y) can
be solved (or closely approximated) by an efficient algorithm that exploits the structure of
the constraints g and basis functions f . This definition covers a broad range of models,
from probabilistic models such as Markov networks and context free grammars to more
unconventional models like weighted graph-cuts and matchings.
12.1.1 Structured maximum margin estimation
Given a sample S = ¦(x
(i)
, y
(i)

m
i=1
, we develop methods for finding parameters w such
that:
arg max
y∈¸
(i)
w
¯
f (x
(i)
, y) ≈ y
(i)
, ∀i,
where ¸
(i)
= ¦y : g(x
(i)
, y) ≤ 0¦.
The naive formulation
1
uses

i

(i)
[ linear constraints, which is generally exponential
in the number of variables in each y
(i)
.
min
1
2
[[w[[
2
s.t. w
¯
f
i
(y
(i)
) ≥ w
¯
f
i
(y) +
i
(y), ∀i, ∀y ∈ ¸
(i)
.
We propose two general approaches that transform the above exponential size QP to an
exactly equivalent polynomial size QP in many important classes of models. These formu-
lations allow us to find globally optimal parameters (with fixed precision) in polynomial
time using standard optimization software. In many models where maximum likelihood
estimation is intractable, we provide exact maximum margin solutions (Ch. 7 and Ch. 10).
1
For simplicity, we omit the slack variables in this summary.
178 CHAPTER 12. CONCLUSIONS AND FUTURE DIRECTIONS
Min-max formulation
We can turn the above problem into an equivalent min-max formulation with i non-linear
max-constraints:
min
1
2
[[w[[
2
s.t. w
¯
f
i
(y
(i)
) ≥ max
y∈¸
(i)
[w
¯
f
i
(y) +
i
(y)], ∀i.
The key to solving the estimation problem above efficiently is the loss-augmented infer-
ence problem max
y∈¸
(i) [w
¯
f
i
(y) +
i
(y)]. Even if max
y∈¸
(i) w
¯
f
i
(y) can be solved in
polynomial time using convex optimization, the form of the loss term
i
(y) is crucial for
the loss-augmented inference to remain tractable. We typically use a natural loss func-
tion which is essentially the Hamming distance between y
(i)
and h(x
(i)
): the number of
variables predicted incorrectly.
We show that if we can express the (loss-augmented) inference as a compact convex
optimization problem (e.g., LP, QP, SDP, etc.), we can embed the maximization inside the
min-max formulation to get a compact convex program equivalent to the naive exponential
formulation. We show that this approach leads to exact polynomial-size formulations for
estimation of low-treewidth Markov networks, associative Markov networks over binary
variables, context-free grammars, bipartite matchings, and many other models.
Certificate formulation
There are several important combinatorial problems which allow polynomial time solu-
tion yet do not have a compact convex optimization formulation. For example, maximum
weight perfect (non-bipartite) matching and spanning tree problems can be expressed as
linear programs with exponentially many constraints, but no polynomial formulation is
known [Bertsimas & Tsitsiklis, 1997; Schrijver, 2003]. Both of these problems, however,
can be solved in polynomial time using combinatorial algorithms. In some cases, though,
we can find a compact certificate of optimality that guarantees that
y
(i)
= arg max
y
[w
¯
f
i
(y) +
i
(y)].
12.1. SUMMARY OF CONTRIBUTIONS 179
For perfect (non-bipartite) matchings, this certificate is a condition that ensures there are
no augmenting alternating cycles (see Ch. 10). We can express this condition by defining
an auxiliary distance function on the nodes an a set of linear constraints that are satisfied if
and only if there are no negative cycles. This simple set of linear constraints scales linearly
with the number of edges in the graph. Similarly, we can derive a compact certificate for
the spanning tree problem.
The certificate formulation relies on the fact that verifying optimality of a solution is
often easier than actually finding one. This observation allows us to apply our framework
to an even broader range of models with combinatorial structure than the min-max formu-
lation.
Maximum margin vs. maximum likelihood
There are several theoretical advantages to our approach in addition to the empirical accu-
racy improvements we have shown experimentally. Because our approach only relies on
using the maximum in the model for prediction, and does not require a normalized dis-
tribution P(y [ x) over all outputs, maximum margin estimation can be tractable when
maximum likelihood is not. For example, to learn a probabilistic model P(y [ x) over
bipartite matchings using maximum likelihood requires computing the normalizing parti-
tion function, which is #P-complete [Valiant, 1979; Garey & Johnson, 1979]. By contrast,
maximum margin estimation can be formulated as a compact QP with linear constraints.
Similar results hold for an important subclass of Markov networks and non-bipartite match-
ings.
In models that are tractable for both maximum likelihood and maximum margin (such
as low-treewidth Markov networks, context free grammars, many other problems in which
inference is solvable by dynamic programming), our approach has an additional advantage.
Because of the hinge-loss, the solutions to the estimation are relatively sparse in the dual
space (as in SVMs), which makes the use of kernels much more efficient. Maximum like-
lihood models with kernels are generally non-sparse and require pruning or greedy support
vector selection methods [Wahba et al., 1993; Zhu & Hastie, 2001; Lafferty et al., 2004;
Altun et al., 2004].
180 CHAPTER 12. CONCLUSIONS AND FUTURE DIRECTIONS
There are, of course, several advantages to maximum likelihood estimation. In appli-
cations where probabilistic confidence information is a must, maximum likelihood is much
more appropriate. Also, in training settings with missing data and hidden variables, proba-
bilistic interpretation permits the use of well-understood algorithms such as EM [Dempster
et al., 1977].
Approximations
In many problems, the maximization problem we are interested in may be NP-hard, for
example, we consider MAP inference in large treewidth Markov networks in Ch. 8, multi-
way cuts in Ch. 7, graph-partitioning in Ch. 11. Many such problems can be written as
integer programs. Relaxations of such integer programs into LPs, QPs or SDPs often pro-
vide excellent approximation algorithms and fit well within our framework, particularly the
min-max formulation. We show empirically that these approximations are very effective in
many applications.
12.1.2 Markov networks: max-margin, associative, relational
The largest portion of the thesis is devoted to novel estimation algorithms, representational
extensions, generalization analysis and experimental validation for Markov networks.
◦ Low-treewidth Markov networks
We use a compact LP for MAP inference in Markov networks with sequence and
other low-treewidth structure to derive an exact, compact, convex learning formu-
lation. The dual formulation allows efficient integration of kernels with graphical
models that leverages rich high-dimensional representations for accurate prediction
in real-world tasks.
◦ Scalable online algorithm
Although our convex formulation is a QP with linear number of variables and con-
straints in the size of the data, for large datasets (millions of examples), it is very
difficult to solve using standard software. We present an efficient algorithm for solv-
ing the estimation problem called Structured SMO. Our online-style algorithm uses
12.1. SUMMARY OF CONTRIBUTIONS 181
inference in the model and analytic updates to solve extremely large estimation prob-
lems.
◦ Generalization analysis
We analyze the theoretical generalization properties of max-margin estimation in
Markov networks and derive a novel margin-based bound for structured prediction.
This is the first bound to address structured error (e.g., proportion of mislabeled
pixels in an image).
◦ Learning associative Markov networks (AMNs)
We define an important subclass of Markov networks that captures positive correla-
tions present in many domains. This class of networks extends the Potts model [Potts,
1952] often used in computer vision and allows exact MAP inference in the case of
binary variables. We show how to express the inference problem using an LP which
is exact for binary networks. As a result, for associative Markov networks over bi-
nary variables, our framework allows exact estimation of networks of arbitrary con-
nectivity and topology, for which likelihood methods are believed to be intractable.
For the non-binary case, we provide an approximation that works well in practice.
We present an AMN-based method for object segmentation from 3D range data. By
constraining the class of Markov networks to AMNs, our models are learned effi-
ciently and, at run-time, can scale up to tens of millions of nodes and edges by using
graph-cut based inference [Kolmogorov & Zabih, 2002].
◦ Representation and learning of relational Markov networks
We introduce relational Markov networks (RMNs), which compactly define tem-
plates for Markov networks for domains with relational structure: objects, attributes,
relations. The graphical structure of an RMN is based on the relational structure of
the domain, and can easily model complex interaction patterns over related entities.
We apply this class of models to classification of hypertext using hyperlink structure
to define relations between webpages. We use a compact approximate MAP LP in
these complex Markov networks, in which exact inference is intractable, to derive an
approximate max-margin formulation.
182 CHAPTER 12. CONCLUSIONS AND FUTURE DIRECTIONS
12.1.3 Broader applications: parsing, matching, clustering
The other large portion of the thesis addresses a range of prediction tasks with very diverse
models: context free grammars for natural language parsing, perfect matchings for disulfide
connectivity in protein structure prediction, graph partitions for clustering documents and
segmenting images.
◦ Learning to parse
We exploit dynamic programming decomposition of context free grammars to derive
a compact max-margin formulation. We build on a recently proposed “unlexicalized”
grammar that allows cubic time parsing and we show how to achieve high-accuracy
parsing (still in cubic time) by exploiting novel kinds of lexical information. We show
experimental evidence of the model’s improved performance over several baseline
models.
◦ Learning to match
We use a combinatorial optimality condition, namely the absence of augmenting al-
ternating cycles, to derive an exact, efficient certificate formulation for learning to
match. We apply our framework to prediction of disulfide connectivity in proteins
using perfect matchings. The algorithmwe propose uses kernels, which makes it pos-
sible to efficiently embed input features in very high-dimensional spaces and achieve
state-of-the-art accuracy.
◦ Learning to cluster
By expressing the correlation clustering problem as a compact LP and SDP, we use
the min-max formulation to learn a parameterized scoring function for clustering. In
contrast to algorithms that learn a metric independently of the algorithm that will
be used to cluster the data, we describe a formulation that tightly integrates metric
learning with the clustering algorithm, tuning one to the other in a joint optimization.
We formulate the approximate learning problem as a compact convex program. Ex-
periments on synthetic and real-world data show the ability of the algorithm to learn
an appropriate clustering metric for a variety of desired clusterings, including email
folder organization and image segmentation.
12.2. EXTENSIONS AND OPEN PROBLEMS 183
12.2 Extensions and open problems
There are several immediate applications, less immediate extensions and open problems for
our estimation framework. We organize these ideas into several sections below, including
further theoretical analysis and new optimization algorithms, novel prediction tasks, and
more general learning settings.
12.2.1 Theoretical analysis and optimization algorithms
◦ Approximation bounds
In several of the intractable models, like multi-class AMNs in Ch. 7 and correlation
clustering in Ch. 11, we used approximate convex programs within the min-max for-
mulation. These approximate inference programs have strong relative error bounds.
An open question is to translate these error bounds on inference into error bounds on
the resulting max-margin formulations.
◦ Generalization bounds with distributional assumptions
In Ch. 5, we presented a bound on the structured error in Markov networks, with-
out any assumption about the distribution of P(y [ x), relying only on the samples
(x
(i)
, y
(i)
) being i.i.d. This distribution-free assumption leads to a worst case analy-
sis, while some assumptions about the approximate decomposition P(y [ x) may be
warranted. For example, for sequential prediction problems, the Markov assumption
of some finite order is reasonable (i.e., given the input and previous k labels, the next
label is independent of the labels more than k in the past). In spatial prediction tasks,
a label variable is independent of the rest given a large enough ball of labels around
it. Similar assumptions may be made for some “degree of separation” in relational
domains. More generally, it would be interesting to exploit such conditional indepen-
dence assumptions or asymptotic bounds on entropy of P(y [ x) to get generalization
guarantees even from a single structured example (x, y).
◦ Problem-specific optimization methods
Although our convex formulations are polynomial in the size of the data, scaling
184 CHAPTER 12. CONCLUSIONS AND FUTURE DIRECTIONS
up to larger datasets will require problem-specific optimization methods. For low-
treewidth Markov networks and context free grammars, we have presented the Struc-
tured SMO algorithm. Another algorithm useful for such models is Exponentiated
Gradient [Bartlett et al., 2004]. Both algorithms rely on dynamic programming de-
compositions. However, models which do not permit such decompositions, such as
graph-cuts, matchings, and many others, create a need for new algorithms that can ef-
fectively use combinatorial optimization as a subroutine to eliminate the dependence
on general-purpose convex solvers.
12.2.2 Novel prediction tasks
◦ Bipartite matchings
Maximum weight bipartite matchings are used in a variety of problems to predict
mappings between sets of items. In machine translation, matchings are used to
map words of the two languages in aligned sentences [Matusov et al., 2004]. In 2D
shape matching, points on two shapes are matched based on their local contour fea-
tures [Belongie et al., 2002]. Our framework provides an exact, efficient alternative
to the maximum likelihood estimation for learning the matching scoring function.
◦ Sequence alignment
In standard pairwise alignment of biological sequences, a string edit distance is used
to determine which portions of the sequences align to each other [Needleman &
Wunsch, 1970; Durbin et al., 1998]. Finding the best alignment involves a dynamic
program that generalizes the longest common subsequence algorithm. Our frame-
work can be applied (just as in context free grammar estimation) to efficiently learn
a more complex edit function that depends on the contextual string features, perhaps
using novel string kernels [Haussler, 1999; Leslie et al., 2002; Lodhi et al., 2000].
◦ Continuous prediction problems
We have addressed estimation of models with discrete output spaces, generalizing
classification models to multivariate, structured classification. Similarly, we can
consider a whole range of problems where the prediction variables are continuous.
12.2. EXTENSIONS AND OPEN PROBLEMS 185
Such problems are a natural generalizations of regression, involving correlated, inter-
constrained real-valued outputs. For example, several recent models of metabolic
flux in yeast use linear programming formulations involving quantities of various
enzymes, with stoichiometric constraints [Varma & Palsson, 1994]. It would be
interesting to use observed equilibria data under different conditions to learn what
“objective” the cell is maximizing. In financial modeling, convex programs are often
used to model portfolio management; for example, Markowitz portfolio optimization
is formulated as a quadratic program which minimizes risk and maximizes expected
return under budget constraints [Markowitz, 1991; Luenberger, 1997]. In this setting,
one could learn a user’s return projection and risk assessment function from observed
portfolio allocations by the user.
These problems are similar to the discrete structured prediction models we have con-
sidered: inference in the model can formulated as a convex optimization problem.
However, there are obstacles to directly applying the min-max or certificate formu-
lations. Details of this are beyond the scope of this thesis, but it suffices to say that
loss-augmented inference using, Hamming distance equivalent, L
1
loss (or L
2
loss),
no longer produces a maximization of a concave objective with convex constraints
since L
1
, L
2
are convex, not concave (it turns out that it is actually possible to use
L

loss). Developing effective loss functions and max-margin formulations for the
continuous setting could provide a novel set of effective models for structured multi-
variate real-valued prediction problems.
12.2.3 More general learning settings
◦ Structure learning
We have focused on the problem of learning parameters of the model (even though
our kernelized models can be considered non-parametric). In the case of Markov
networks, especially in spatial and relational domains, there is a wealth of possible
structures (cliques in the network) one can use to model a problem. It is particularly
interesting to explore the problem of inducing these cliques automatically from data.
The standard method of greedy stepwise selection followed by re-estimation is very
186 CHAPTER 12. CONCLUSIONS AND FUTURE DIRECTIONS
expensive in general networks [Della Pietra et al., 1997; Bach & Jordan, 2001].
Recent work on selecting input features in Markov networks (or CRFs) uses several
approximations to learn efficiently with millions of candidate features [McCallum,
2003]. However, clique selection is still relatively unexplored. It is possible that
AMNs, by restricting the network to be tractable under any structure, may permit
more efficient clique selection methods.
◦ Semi-supervised learning
Throughout the thesis we have assumed completely labeled data. This assumption
often limits us to relatively small training sets where data has been carefully anno-
tated, while much of the easily accessible data is not at all or suitably labeled. There
are several more general settings we would like to extend our framework.
The simplest setting is a mix of labeled and unlabeled examples, where a small su-
pervised dataset is augmented by a large unsupervised one. There has been much
research in this setting for classification [Blum & Mitchell, 1998; Nigam et al., 2000;
Chapelle et al., 2002; Szummer & Jaakkola, 2001; Zhu et al., 2003; Corduneanu &
Jaakkola, 2003]. Although most of this work has been done in a probabilistic set-
ting, the principle of regularizing (discouraging) decision boundaries near densely
clustered inputs could be applicable to our structured setting.
A more complex and rich setting involves presence of hidden variables in each ex-
ample. For example, in machine translation, word correspondences between pairs of
sentences are usually not manually annotated (at least not on a large scale). These
correspondence variables can be treated as hidden variables. Similarly, in handwrit-
ing recognition, we may not have each letter segmented out but instead just get a
word or sentence as a label for the entire image. This setting has been studied mainly
in the probabilistic, generative models often using the EMalgorithm[Dempster et al.,
1977; Cowell et al., 1999]. Discriminative methods have been explored far less. Es-
pecially in the case of combinatorial structures, extensions of our framework allow
opportunities for problem-specific convex approximations to be exploited.
12.3. FUTURE 187
12.3 Future
We have presented a supervised learning framework for a large class of prediction mod-
els with rich and interesting structure. Our approach has several theoretical and practical
advantages over standard probabilistic models and estimation methods for structured pre-
diction. We hope that continued research in this framework will help tackle evermore
sophisticated prediction problems in the future.
Appendix A
Proofs and derivations
A.1 Proof of Theorem 5.5.1
The proof of Theorem 5.5.1 uses the covering number bounds of Zhang [2002] (in the
Data-Dependent Structural Risk Minimization framework [Shawe-Taylor et al., 1998].)
Zhang provides generalization guarantees for linear binary classifiers of the form h
w
(x) =
sgn(w
¯
x). His analysis is based on the upper bounds on the covering number for the class
of linear functions T
L
(w, z) = w
¯
z where the norms of the vectors w and z are bounded.
We reproduce the relevant definitions and theorems from Zhang [2002] here to highlight
the necessary extensions for structured classification.
The covering number is a key quantity in measuring function complexity. Intuitively,
the covering number of an infinite class of functions (e.g. parameterized by a set of weights
w) is the number of vectors necessary to approximate the values of any function in the class
on a sample. Margin-based analysis of generalization error uses the margin achieved by a
classifier on the training set to approximate the original function class of the classifier by
a finite covering with precision that depends on the margin. Here, we will only define the
∞-norm covering number.
188
A.1. PROOF OF THEOREM 5.5.1 189
A.1.1 Binary classification
In binary classification, we are given a sample S = ¦x
(i)
, y
(i)
¦
m
i=1
, from distribution D over
A ¸, where A = IR
n
and ¸ is mapped to ±1, so we can fold x and y into z = yx.
Definition A.1.1 (Covering Number) Let 1 = ¦v
(1)
, . . . , v
(r)
¦, where v
(j)
∈ IR
m
, be a
covering of a function class T(w, S) with -precision under the metric ρ, if for all w there
exists a v
(j)
such that for each data sample z
(i)
∈ S:
ρ(v
(j)
i
, T(w, z
(i)
)) ≤ .
The covering number of a sample S is the size of the smallest covering: ^

(T, ρ, , S) =
inf [1[ s.t. 1 is a covering of T(w, S). We also define the covering number for any sample
of size m: ^

(T, ρ, , m) = sup
S: [S[=m
^

(T, ρ, , S).
When the norms of w and z are bounded, we have the following upper bound on the
covering number of linear functions under the linear metric ρ
L
(v, v
t
) = [v −v
t
[.
Theorem A.1.2 (Theorem 4 from Zhang [2002]) If |w|
2
≤ a and |z|
2
≤ b, then ∀
> 0,
log
2
^

(T
L
, ρ
L
, , m) ≤ 36
a
2
b
2

2
log
2
(2 ¸4ab/ + 2| m + 1) .
In order to use the classifier’s margin to bound its expected loss, the bounds below use
a stricter, margin-based loss on the training sample that measures the worst loss achieved
by the approximate covering based on this margin. Let f : IR → [0, 1] be a loss function.
In binary classification, we let f(v) = 1I(v ≤ 0) be the step function, so that 0-1 loss of
sgn(w
¯
x) is f(T
L
(w, z)). The next theorem bounds the expected f loss in terms of the
γ-margin loss, f
γ
(v) = sup
ρ(v,v

)<2γ
f(v
t
), on the training sample. For 0-1 loss and linear
metric ρ
L
, the corresponding γ-margin loss is f
γ
(v) = 1I(v ≤ 2γ).
Theorem A.1.3 (Corollary 1 from Zhang [2002]) Let f : IR → [0, 1] be a loss function
and f
γ
(v) = sup
ρ(v,v

)<2γ
f(v
t
) be the γ-margin loss for a metric ρ. Let γ
1
> γ
2
> . . . be
190 APPENDIX A. PROOFS AND DERIVATIONS
a decreasing sequence of parameters, and p
i
be a sequence of positive numbers such that


i=1
p
i
= 1, then for all δ > 0, with probability of at least 1 −δ over data:
E
D
[f(T(w, z))] ≤ E
S
[f
γ
(T(w, z))] +
¸
32
m
_
ln 4^

(T, ρ, γ
i
, S) + ln
1
p
i
δ
_
for all w and γ, where for each fixed γ, we use i to denote the smallest index s.t. γ
i
≤ γ.
A.1.2 Structured classification
We will extend this framework to bound the average per-label loss
H
(y)/L for structured
classification by defining an appropriate loss f and a function class T (as well as a metric
ρ) such that f(T) computes average per-label loss and f
γ
(T) provides a suitable γ-margin
loss. We will bound the corresponding covering number by building on the bound in The-
orem A.1.2.
We can no longer simply fold x and y, since y is a vector, so we let z = (x, y). In
order for our loss function to compute average per-label loss, it is convenient to make our
function class vector-valued (instead of scalar-valued as above). We define a new function
class T
M
(w, z), which is a vector of minimum values of w
¯
∆f
i
(y) for each error level

H
(y) from 1 to L as described below.
Definition A.1.4 (dth-error-level function) The dth-error-level function M
d
(w, z) for d ∈
¦1, . . . , L¦ is given by:
M
d
(w, z) = min
y:
H
(y)=d
w
¯
∆f
i
(y).
Definition A.1.5 (Multi-error-level function class) The multi-error-level function class T
M
(w, z)
is given by:
T
M
(w, z) = (M
1
(w, z), . . . , M
d
(w, z), . . . , M
L
(w, z)) .
We can now compute the average per-label loss from T
M
(w, z) by defining an appropriate
loss function f
M
.
A.1. PROOF OF THEOREM 5.5.1 191
Definition A.1.6 (Average per-label loss) The average per-label loss f
M
: IR
L
→[0, 1] is
given by:
f
M
(v) =
1
L
arg min
d:v
d
≤0
v
d
,
where in case ∀d, v
d
> 0, we define arg min
d:v
d
≤0
v
d
≡ 0.
With the above definitions, we have an upper bound on the average per-label loss
f
M
(T
M
(w, z)) =
1
L
arg min
d:M
d
(w,z)≤0
M
d
(w, z) ≥
1
L

H
_
arg max
y
w
¯
f
i
(y)
_
.
Note that the case ∀d, M
d
(w, z) > 0 corresponds to the classifier making no mistakes:
arg max
y
w
¯
f
i
(y) = y. This upper bound is tight if y = arg max
y
w
¯
f (x, y
t
), Other-
wise, it is adversarial: it picks from all y
t
which are better (w
¯
f (y) ≤ w
¯
f (y
t
)), one that
maximizes the Hamming distance from y.
We now need to define an appropriate metric ρ that in turn defines γ-margin loss for
structured classification. Since the margin of the hypothesis grows with the number of
mistakes, our metric can become “looser” with the number of mistakes, as there is more
room for error.
Definition A.1.7 (Multi-error-level metric) Let the multi-error-level metric ρ
M
: IR
L

IR
L
→IR for a vector in IR
L
be given by:
ρ
M
(v, v
t
) = max
d
[v
d
−v
t
d
[
d
.
We now define the corresponding γ-margin loss using the new metric:
Definition A.1.8 (γ-margin average per-label loss) The γ-margin average per-label loss
f
γ
M
: IR
L
→[0, 1] is given by:
f
γ
M
(v) = sup
ρ
M
(v,v

)≤2γ
f
M
(v
t
).
Combining the two definitions, we get:
f
γ
M
(T
M
(w, z)) = sup
v:[v
d
−M
d
(w,z)[≤2dγ
1
L
arg min
d:v
d
≤0
v
d
.
192 APPENDIX A. PROOFS AND DERIVATIONS
We also define the corresponding covering number for our vector-valued function class:
Definition A.1.9 (Multi-error-level covering number) Let 1 = ¦V
(1)
, . . . , V
(r)
¦, where
V
(j)
= (V
(j)
1
, . . . , V
(j)
i
, . . . , V
(j)
m
) and V
(j)
i
∈ IR
L
, be a covering of T
M
(w, S), with -
precision under the metric ρ
M
, if for all w there exists a V
(j)
such that for each data
sample z
(i)
∈ S:
ρ
M
(V
(j)
i
, T
M
(w, z
(i)
)) ≤ .
The covering number of a sample S is the size of the smallest covering: ^

(T
M
, ρ
M
, , S) =
inf [1[ s.t. 1 is a covering of T
M
(w, S). We also define
^

(T
M
, ρ
M
, , m) = sup
S: [S[=m
^

(T
M
, ρ
M
, , S).
We provide a bound on the covering number of our new function class in terms of a
covering number for the linear function class. Recall that N
c
is the maximum number of
cliques in ((x), V
c
is the maximum number of values in a clique [Y
c
[, q is the maximum
number of cliques that have a variable in common, and R
c
is an upper-bound on the 2-norm
of clique basis functions. Consider a first-order sequence model as an example, with L as
the maximum length, and V the number of values a variable takes. Then N
c
= 2L−1 since
we have L node cliques and L −1 edge cliques; V
c
= V
2
because of the edge cliques; and
q = 3 since nodes in the middle of the sequence participate in 3 cliques: previous-current
edge clique, node clique, and current-next edge clique.
Lemma A.1.10 (Bound on multi-error-level covering number)
^

(T
M
, ρ
M
, q, m) ≤ ^

(T
L
, ρ
L
, , mN
c
(V
c
−1)).
Proof: We will show that ^

(T
M
, ρ
M
, q, S) ≤ ^

(T
L
, ρ
L
, , S
t
) for any sample S
of size m, where we construct the sample S
t
of size mN
c
(V
c
− 1) in order to cover the
clique potentials as described below. Note that this is sufficient since ^

(T
L
, ρ
L
, , S
t
) ≤
^

(T
L
, ρ
L
, , mN
c
(V
c
−1)), by definition, so
^

(T
M
, ρ
M
, q, m) = sup
S:[S[=m
^

(T
M
, ρ
M
, q, S) ≤ ^

(T
L
, ρ
L
, , mN
c
(V
c
−1)).
A.1. PROOF OF THEOREM 5.5.1 193
The construction of S
t
below is inspired by the proof technique in Collins [2001],
but the key difference is that our construction is linear in the number of cliques N
c
and
exponential in the number of label variables per clique, while his is exponential in the total
number of label variables per example. This reduction in size comes about because our
covering approximates the values of clique potentials w
¯
∆f
i,c
(y
c
) for each clique c and
clique assignment y
c
as opposed to the values of entire assignments w
¯
∆f
i
(y).
For each sample z ∈ S, we create N
c
(V
c
− 1) samples ∆f
i,c
(y
c
), one for each clique
c and each assignment y
c
,= y
(i)
c
. We construct a set of vectors 1 = ¦v
(1)
, . . . , v
(r)
¦,
where v
(j)
∈ IR
mN
c
(V
c
−1)
. The component of v
(j)
corresponding to the sample z
(i)
and the
assignment y
c
to the labels of the clique c will be denoted by v
(j)
i,c
(y
c
). For convenience,
we define v
(j)
i,c
(y
(i)
c
) = 0 for correct label assignments, as ∆f
i,c
(y
(i)
c
) = 0. To make 1 an
∞-norm covering of T
L
(w, S
t
) under ρ
L
, we require that for any w there exists a v
(j)
∈ 1
such that for each sample z
(i)
:
[v
(j)
i,c
(y
c
) −w
¯
∆f
i,c
(y
c
)[ ≤ ; ∀c ∈ (
(i)
, ∀y
c
. (A.1)
By Definition A.1.1, the number of vectors in 1 is given by r = ^

(T
L
, ρ
L
, , mN
c
(V
c
−1)).
We can now use 1 to construct a covering 1 = ¦V
(1)
, . . . , V
(r)
¦, where
V
(j)
= (V
(j)
1
, . . . , V
(j)
i
, . . . , V
(j)
m
)
and V
(j)
i
∈ IR
L
, for our multi-error-level function T
M
. Let v
(j)
i
(y) =

c
v
(j)
i,c
(y
c
), and
M
d
(v
(j)
i
, z
(i)
) = min
y:
H
i
(y)=d
v
(j)
i
(y), then
V
(j)
i
= (M
1
(v
(j)
, z
(i)
), . . . , M
d
(v
(j)
, z
(i)
), . . . , M
L
(v
(j)
, z
(i)
)) . (A.2)
Note that v
(j)
i,c
(y
c
) is zero for all cliques c for which the assignment is correct: y
c
= y
(i)
c
.
Thus for an assignment y with d mistakes, at most dq v
(j)
i,c
(y
c
) will be non-zero, as each
label can appear in at most q cliques. By combining this fact with Eq. (A.1), we obtain:
¸
¸
¸v
(j)
i
(y) −w
¯
∆f
i
(y)
¸
¸
¸ ≤ dq, ∀i, ∀y :
H
i
(y) = d. (A.3)
194 APPENDIX A. PROOFS AND DERIVATIONS
We conclude the proof by showing that 1 is a covering of T
M
under ρ
M
: For each w,
pick V
(j)
∈ 1 such that the corresponding v
(j)
∈ 1 satisfies the condition in Eq. (A.1). We
must now bound:
ρ
M
(V
(j)
i
, T
M
(w, z
(i)
)) = max
d
[ min
y:
H
i
(y)=d
v
(j)
i
(y) −min
y:
H
i
(y)=d
w
¯
∆f
i
(y)[
d
.
Fix any i. Let y
v
d
= arg min
y:
H
i
(y)=d
v
(j)
i
(y) and y
w
d
= arg min
y:
H
i
(y)=d
w
¯
∆f
i
(y). Con-
sider the case where v
(j)
i
(y
v
d
) ≥ w
¯
∆f
i
(y
w
d
) (the reverse case is analogous), we must
prove that:
v
(j)
i
(y
v
d
) −w
¯
∆f
i
(y
w
d
) ≤ v
(j)
i
(y
w
d
) −w
¯
∆f
i
(y
w
d
) ≤ dq ; (A.4)
where the first step follows from definition of y
v
d
, since v
(j)
i
(y
v
d
) ≤ v
(j)
i
(y
w
d
). The last step
is a direct consequence of Eq. (A.3). Hence ρ
M
(V
(j)
i
, T
M
(w, z
(i)
)) ≤ q.
Lemma A.1.11 (Numeric bound on multi-error-level covering number)
log
2
^

(T
M
, ρ
M
, , m) ≤ 36
R
2
c
|w|
2
2
q
2

2
log
2
_
1 + 2
_
4
R
c
|w|
2
q

+ 2
_
mN
c
(V
c
−1)
_
.
Proof: Substitute Theorem A.1.2 into Lemma A.1.10.
Theorem A.1.12 (Multi-label analog of Theorem A.1.3) Let f
M
and f
γ
M
(v) be as defined
above. Let γ
1
> γ
2
> . . . be a decreasing sequence of parameters, and p
i
be a sequence
of positive numbers such that


i=1
p
i
= 1, then for all δ > 0, with probability of at least
1 −δ over data:
E
z
f
M
(T
M
(w, z)) ≤ E
S
f
γ
M
(T
M
(w, z)) +
¸
32
m
_
ln 4^

(T
M
, ρ
M
, γ
i
, S) + ln
1
p
i
δ
_
for all w and γ, where for each fixed γ, we use i to denote the smallest index s.t. γ
i
≤ γ.
Proof: Similar to the proof of Zhang’s Theorem 2 and Corollary 1 [Zhang, 2002] where
in Step 3 (derandomization) we substitute the vector-valued T
M
and the metric ρ
M
.
Theorem 5.5.1 follows from above theorem with γ
i
= R
c
|w|
2
/2
i
and p
i
= 1/2
i
using an
A.2. AMN PROOFS AND DERIVATIONS 195
argument identical to the proof of Theorem 6 in Zhang [2002].
A.2 AMN proofs and derivations
In this appendix, we present proofs of the LP inference properties and derivations of the
factored primal and dual max-margin formulation from Ch. 7. Recall that the LP relaxation
for finding the optimal max
y
g(y) is:
max

v∈1
K

k=1
µ
v
(k)g
v
(k) +

c∈(\1
K

k=1
µ
c
(k)g
c
(k) (A.5)
s.t. µ
c
(k) ≥ 0, ∀c ∈ (, k;
K

k=1
µ
v
(k) = 1, ∀v ∈ 1;
µ
c
(k) ≤ µ
v
(k), ∀c ∈ ( ¸ 1, v ∈ c, k.
A.2.1 Binary AMNs
Proof (For Theorem 7.2.1) Consider any fractional, feasible µ. We show that we can con-
struct a new feasible assignment µ
t
which increases the objective (or leaves it unchanged)
and furthermore has fewer fractional entries.
Since g
c
(k) ≥ 0, we can assume that µ
c
(k) = min
v∈c
µ
v
(k); otherwise we could in-
crease the objective by increasing µ
c
(k). We construct an assignment µ
t
from µ by leaving
integral values unchanged and uniformly shifting fractional values by λ:
µ
t
v
(1) = µ
v
(1) −λ1I(0 < µ
v
(1) < 1), µ
t
v
(2) = µ
v
(2) + λ1I(0 < µ
v
(2) < 1),
µ
t
c
(1) = µ
c
(1) −λ1I(0 < µ
c
(1) < 1), µ
t
c
(2) = µ
c
(2) + λ1I(0 < µ
c
(2) < 1).
Now consider the smallest fractional µ
v
(k), λ(k) = min
v : µ
v
(k)>0
y
v
(k) for k = 1, 2.
Note that if λ = λ(1) or λ = −λ(2), µ
t
will have at least one more integral µ
t
v
(k) than µ.
Thus if we can show that the update results in a feasible and better scoring assignment, we
can apply it repeatedly to get an optimal integer solution. To show that µ
t
is feasible, we
need µ
t
v
(1) + µ
t
v
(2) = 1, µ
t
v
(k) ≥ 0 and µ
t
c
(k) = min
i∈c
µ
t
v
(k).
196 APPENDIX A. PROOFS AND DERIVATIONS
First, we show that µ
t
v
(1) + µ
t
v
(2) = 1.
µ
t
v
(1) + µ
t
v
(2) = µ
v
(1) −λ1I(0 < µ
v
(1) < 1) + µ
v
(2) + λ1I(0 < µ
v
(2) < 1)
= µ
v
(1) + µ
v
(2) = 1.
Above we used the fact that if µ
v
(1) is fractional, so is µ
v
(2), since µ
v
(1) + µ
v
(2) = 1.
To show that µ
t
v
(k) ≥ 0, we prove min
v
µ
t
v
(k) = 0.
min
v
µ
t
v
(k) = min
v
_
µ
v
(k) −( min
i:µ
v
(k)>0
µ
v
(k))1I(0 < µ
v
(k) < 1)
_
= min
_
min
i
µ
v
(k), min
i:µ
v
(k)>0
_
µ
v
(k) − min
i:µ
v
(k)>0
µ
v
(k)
__
= 0.
Lastly, we show µ
t
c
(k) = min
i∈c
µ
t
v
(k).
µ
t
c
(1) = µ
c
(1) −λ1I(0 < µ
c
(1) < 1)
= (min
i∈c
µ
v
(1)) −λ1I(0 < min
i∈c
µ
v
(1) < 1) = min
i∈c
µ
t
v
(1);
µ
t
c
(2) = µ
c
(2) + λ1I(0 < µ
c
(1) < 1)
= (min
i∈c
µ
v
(2)) + λ1I(0 < min
i∈c
µ
v
(2) < 1) = min
i∈c
µ
t
v
(2).
We have established that the new µ
t
are feasible, and it remains to show that we can
improve the objective. We can show that the change in the objective is always λD for some
constant D that depends only on µ and g. This implies that one of the two cases, λ = λ(1)
or λ = −λ(2), will necessarily increase the objective (or leave it unchanged). The change
in the objective is:

v∈1

k=1,2

t
v
(k) −µ
v
(k)]g
v
(k) +

c∈(\1

k=1,2

t
c
(k) −µ
c
(k)]g
c
(k)
= λ
_
_

v∈1
[D
v
(1) −D
v
(2)] +

c∈(\1
[D
c
(1) −D
c
(2)]
_
_
= λD
D
v
(k) = g
v
(k)1I(0 < µ
v
(k) < 1), D
c
(k) = g
c
(k)1I(0 < µ
c
(k) < 1).
Hence the new assignment µ
t
is feasible, does not decrease the objective function, and
A.2. AMN PROOFS AND DERIVATIONS 197
has strictly fewer fractional entries.
A.2.2 Multi-class AMNs
For K > 2, we use the randomized rounding procedure of Kleinberg and Tardos [1999]
to produce an integer solution for the linear relaxation, losing at most a factor of m =
max
c∈(
[c[ in the objective function. The basic idea of the rounding procedure is to treat
µ
v
(k) as probabilities and assign labels according to these probabilities in phases. In each
phase, we pick a label k, uniformly at random, and a threshold α ∈ [0, 1] uniformly at
random. For each node i which has not yet been assigned a label, we assign the label k
if µ
v
(k) ≥ α. The procedure terminates when all nodes have been assigned a label. Our
analysis closely follows that of Kleinberg and Tardos [1999].
Lemma A.2.1 The probability that a node i is assigned label k by the randomized proce-
dure is µ
v
(k).
Proof The probability that an unassigned node is assigned label k during one phase is
1
K
µ
v
(k), which is proportional to µ
v
(k). By symmetry, the probability that a node is as-
signed label k over all phases is exactly µ
v
(k).
Lemma A.2.2 The probability that all nodes in a clique c are assigned label k by the
procedure is at least
1
[c[
µ
c
(k).
Proof For a single phase, the probability that all nodes in a clique c are assigned label k if
none of the nodes were previously assigned is
1
K
min
i∈c
µ
v
(k) =
1
K
µ
c
(k). The probability
that at least one of the nodes will be assigned label k in a phase is
1
K
(max
i∈c
µ
v
(k)). The
probability that none of the nodes in the clique will be assigned any label in one phase is
1 −
1
K

K
k=1
max
i∈c
µ
v
(k).
Nodes in the clique c will be assigned label k by the procedure if they are assigned label
k in one phase. (They can also be assigned label k as a result of several phases, but we can
ignore this possibility for the purposes of the lower bound.) The probability that all the
198 APPENDIX A. PROOFS AND DERIVATIONS
nodes in c will be assigned label k by the procedure in a single phase is:

j=1
1
K
µ
c
(k)
_
1 −
1
K
K

k=1
max
i∈c
µ
v
(k)
_
j−1
=
µ
c
(k)

K
k=1
max
i∈c
µ
v
(k)

µ
c
(k)

K
k=1

i∈c
µ
v
(k)
=
µ
c
(k)

i∈c

K
k=1
µ
v
(k)
=
µ
c
(k)
[c[
.
Above, we first used the fact that for d < 1,


i=0
d
i
=
1
1−d
, and then upper-bounded
the max of the set of positive µ
v
(k)’s by their sum.
Theorem A.2.3 The expected cost of the assignment found by the randomized procedure
given a solution µ to the linear program in Eq. (A.5) is at least

v∈1

K
k=1
g
v
(k)µ
v
(k) +

c∈(\1
1
[c[

K
k=1
g
c
(k)µ
k
c
.
Proof This is immediate from the previous two lemmas.
The only difference between the expected cost of the rounded solution and the (non-
integer) optimal solution is the
1
[c[
factor in the second term. By picking m = max
c∈(
[c[, we
have that the rounded solution is at most m times worse than the optimal solution produced
by the LP of Eq. (A.5).
We can also derandomize this procedure to get a deterministic algorithm with the same
guarantees, using the method of conditional probabilities, similar in spirit to the approach
of Kleinberg and Tardos [1999].
Note that the approximation factor of m applies, in fact, only to the clique poten-
tials. Thus, if we compare the log-probability of the optimal MAP solution and the log-
probability of the assignment produced by this randomized rounding procedure, the terms
corresponding to the log-partition-function and the node potentials are identical. We obtain
an additive error (in log-probability space) only for the clique potentials. As node poten-
tials are often larger in magnitude than clique potentials, the fact that we incur no loss
proportional to node potentials is likely to lead to smaller errors in practice. Along similar
lines, we note that the constant factor approximation is smaller for smaller cliques; again,
we observe, the potentials associated with large cliques are typically smaller in magnitude,
reducing further the actual error in practice.
A.2. AMN PROOFS AND DERIVATIONS 199
A.2.3 Derivation of the factored primal and dual max-margin QP
Using Assumptions 7.4.1 and 7.4.4, we have the dual of the LP used to represent the
interior max subproblem max
y
w
¯
f
i
(y) +
i
(y) in Eq. (3.2):
min

v∈1
ξ
i,v
(A.6)
s.t. −w
¯
f
i,v
(k) −

c⊃v
m
i,c,v
(k) ≥
i,v
(k) −ξ
i,v
, ∀i, v ∈ 1
(i)
, k;
−¨ w
¯
¨
f
i,c
(k) +

v∈c
m
i,c,v
(k) ≥
i,c
(k), ∀i, c ∈ (
(i)
¸ 1
(i)
, k;
m
i,c,v
(k) ≥ 0, ∀i, c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k;
where f
i,c
(k) = f
i,c
(k, . . . , k) and
i,c
(k) =
i,c
(k, . . . , k). In the dual, we have a variable
ξ
i,v
for each normalization constraint in Eq. (7.1) and variables m
i,c,v
(k) for each of the
inequality constraints.
Substituting this dual into Eq. (5.1), we obtain:
min
1
2
[[w[[
2
+ C

i
ξ
i
(A.7)
s.t. w
¯
f
i
(y
(i)
) + ξ
i

v∈1
(i)
ξ
i,v
, ∀i;
−w
¯
f
i,v
(k) −

c⊃v
m
i,c,v
(k) ≥
i,v
(k) −ξ
i,v
, ∀i, v ∈ 1
(i)
, k;
−¨ w
¯
¨
f
i,c
(k) +

v∈c
m
i,c,v
(k) ≥
i,c
(k), ∀i, c ∈ (
(i)
¸ 1
(i)
, k;
m
i,c,v
(k) ≥ 0, ∀i, c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k;
¨ w ≥ 0.
Now let ξ
i,v
= ξ
t
i,v
+ w
¯
f
i,v
(y
(i)
v
) +

c⊃v
¨ w
¯
¨
f
i,c
(y
(i)
c
)/[c[ and m
i,c,v
(k) = m
t
i,c,v
(k) +
¨ w
¯
¨
f
i,c
(y
(i)
c
)/[c[. Re-expressing the above QP in terms of these new variables, we get:
200 APPENDIX A. PROOFS AND DERIVATIONS
min
1
2
[[w[[
2
+ C

i
ξ
i
(A.8)
s.t. ξ
i

v∈1
(i)
ξ
t
i,v
, ∀i;
w
¯
∆f
i,v
(k) −

c⊃v
m
t
i,c,v
(k) ≥
i,v
(k) −ξ
t
i,v
, ∀i, v ∈ 1
(i)
, k;
¨ w
¯

¨
f
i,c
(k) +

v∈c
m
t
i,c,v
(k) ≥
i,c
(k), ∀i, c ∈ (
(i)
¸ 1
(i)
, k;
m
t
i,c,v
(k) ≥ −¨ w
¯
¨
f
i,c
(y
(i)
c
)/[c[, ∀i, c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k;
¨ w ≥ 0.
Since ξ
i
=

i,v∈1
(i)
ξ
t
i,v
at the optimum, we can eliminate ξ
i
and the corresponding set
of constraints to get the formulation in Eq. (7.4), repeated here for reference:
min
1
2
[[w[[
2
+ C

i,v∈1
(i)
ξ
i,v
(A.9)
s.t. w
¯
∆f
i,v
(k) −

c⊃v
m
i,c,v
(k) ≥
i,v
(k) −ξ
i,v
, ∀i, v ∈ 1
(i)
, k;
¨ w
¯

¨
f
i,c
(k) +

v∈c
m
i,c,v
(k) ≥
i,c
(k), ∀i, c ∈ (
(i)
¸ 1
(i)
, k;
m
i,c,v
(k) ≥ −¨ w
¯
¨
f
i,c
(y
(i)
c
)/[c[, ∀i, c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k;
¨ w ≥ 0.
A.2. AMN PROOFS AND DERIVATIONS 201
Now the dual of Eq. (A.9) is given by:
max

i,c∈(
(i)
, k
µ
i,c
(k)
i,c
(k) −
1
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,v∈1
(i)
, k
µ
i,v
(k)∆
˙
f
i,v
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
(A.10)

1
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¨ τ +

i,c∈(
(i)
\1
(i)
,v∈c, k
λ
i,c,v
(k)
¨
f
i,c
(y
(i)
c
)/[c[ +

i,c∈(
(i)
, k
µ
i,c
(k)∆
¨
f
i,c
(k)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
s.t. µ
i,c
(k) ≥ 0, ∀i, ∀c ∈ (
(i)
, k;
K

k=1
µ
i,v
(k) = C, ∀i, ∀v ∈ 1
(i)
;
µ
i,c
(k) −µ
i,v
(k) = λ
i,c,v
(k), ∀i, ∀c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k;
λ
i,c,v
(k) ≥ 0 ∀i, ∀c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k,
¨ τ ≥ 0.
In this dual, µ correspond to the first two sets of constraints, while λ and ¨ τ correspond
to third and fourth set of constraints. Using the substitution
¨ ν = ¨ τ +

i,c∈(
(i)
\1
(i)
,v∈c, k
λ
i,c,v
(k)
¨
f
i,c
(y
(i)
c
)/[c[
and the fact that λ
i,c,v
(k) ≥ 0 and
¨
f
i,c
(y
(i)
c
) ≥ 0, we can eliminate λ and ¨ τ, as well as divide
µ’s by C, and re-express the above QP as:
max

i,c∈(
(i)
, k
µ
i,c
(k)
i,c
(k) −
1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

i,v∈1
(i)
, k
µ
i,v
(k)∆
˙
f
i,v
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2

1
2
C
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¨ ν +

i,c∈(
(i)
, k
µ
i,c
(k)∆
¨
f
i,c
(k)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
2
s.t. µ
i,c
(k) ≥ 0, ∀i, ∀c ∈ (
(i)
, k;
K

k=1
µ
i,v
(k) = 1, ∀i, ∀v ∈ 1
(i)
;
µ
i,c
(k) ≤ µ
i,v
(k), ∀i, ∀c ∈ (
(i)
¸ 1
(i)
, v ∈ c, k; ¨ ν ≥ 0.
Bibliography
Abbeel, P., & Ng, A. Y. (2004). Apprenticeship learning via inverse reinforcement learning.
Proc. ICML.
Ahuja, R., & Orlin, J. (2001). Inverse optimization, Part I: Linear programming and general
problem. Operations Research, 35, 771–783.
Altschul, S., Madden, T., Schaffer, A., Zhang, A., Miller, W., & Lipman (1997). Gapped
BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleid
Acids Res., 25, 3389–3402.
Altun, Y., Smola, A., & Hofmann, T. (2004). Exponential families for conditional random
fields. Proc. UAI.
Altun, Y., Tsochantaridis, I., & Hofmann, T. (2003). Hidden markov support vector ma-
chines. Proc. ICML.
Bach, F., & Jordan, M. (2001). Thin junction trees. NIPS.
Bach, F., & Jordan, M. (2003). Learning spectral clustering. NIPS.
Baeza-Yates, R., & Ribeiro-Neto, B. (1999). Modern information retrieval. Addison-
Wesley-Longman.
Bairoch, A., & Apweiler, R. (2000). The Swiss-Prot protein sequence database and its
supplement trembl. Nucleic Acids Res., 28, 45–48.
Baldi, P., Cheng, J., & Vullo, A. (2004). Large-scale prediction of disulphide bond connec-
tivity. Proc. NIPS.
202
BIBLIOGRAPHY 203
Baldi, P., & Pollastri, G. (2003). The principled design of large-scale recursive neural
network architectures dag-rnns and the protein structure prediction problem. Journal of
Machine Learning Research, 4, 575–602.
Bansal, N., Blum, A., & Chawla, S. (2002). Correlation clustering. FOCS.
Bar-Hillel, A., Hertz, T., Shental, N., & Weinshall, D. (2003). Learning distance functions
using equivalence relations. Proc. ICML.
Bartlett, P., Collins, M., Taskar, B., & McAllester, D. (2004). Exponentiated gradient
algorithms for large-margin structured classification. NIPS.
Bartlett, P. L., Jordan, M. I., & McAuliffe, J. D. (2003). Convexity, classification, and risk
bounds (Technical Report 638). Department of Statistics, U.C. Berkeley.
Belongie, S., Malik, J., & Puzicha, J. (2002). Shape matching and object recognition using
shape contexts. IEEE Trans. Pattern Anal. Mach. Intell., 24.
Berman, H., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T., Weissig, H., Shindyalov, I., &
Bourne, P. (2000). The protein data bank.
Bertsekas, D. (1999). Nonlinear programming. Belmont, MA: Athena Scientific.
Bertsimas, D., & Tsitsiklis, J. (1997). Introduction to linear programming. Athena Scien-
tific.
Besag, J. E. (1986). On the statistical analysis of dirty pictures. Journal of the Royal
Statistical Society B, 48.
Bishop, C. (1995). Neural networks for pattern recognition. Oxford, UK: Oxford Univer-
sity Press.
Blum, A., & Mitchell, T. (1998). Combining labeled and unlabeled data with co-training.
Proc. COLT.
Bouman, C., & Shapiro, M. (1994). A multiscale random field model for bayesian image
segmentation. IP, 3(2).
204 BIBLIOGRAPHY
Boyd, S., & Vandenberghe, L. (2004). Convex optimization. Cambridge University Press.
Boykov, Y., & Kolmogorov, V. (2004). An experimental comparison of min-cut/max-flow
algorithms for energy minimization in computer vision. PAMI.
Boykov, Y., Veksler, O., & Zabih, R. (1999a). Fast approximate energy minimization via
graph cuts. ICCV.
Boykov, Y., Veksler, O., & Zabih, R. (1999b). Markov random fields with efficient approx-
imations. CVPR.
Burton, D., & Toint, P. L. (1992). On an instance of the inverse shortest paths problem.
Mathematical Programming, 53, 45–61.
Ceroni, A., Frasconi, P., Passerini, A., & Vullo, A. (2003). Predicting the disulfide bond-
ing state of cysteines with combinations of kernel machines. Journal of VLSI Signal
Processing, 35, 287–295.
Chakrabarti, S., Dom, B., & Indyk, P. (1998). Enhanced hypertext categorization using
hyperlinks. SIGMOD.
Chapelle, O., Weston, J., & Schoelkopf, B. (2002). Cluster kernels for semi-supervised
learning. Proc. NIPS.
Charikar, M., Guruswami, V., & Wirth, A. (2003). Clustering with qualitative information.
FOCS.
Charniak, E. (1993). Statistical language learning. MIT Press.
Clark, S., & Curran, J. R. (2004). Parsing the wsj using ccg and log-linear models. Pro-
ceedings of the 42nd Annual Meeting of the Association for Computational Linguistics
(ACL ’04).
Collins, M. (1999). Head-driven statistical models for natural language parsing. Doctoral
dissertation, University of Pennsylvania.
BIBLIOGRAPHY 205
Collins, M. (2000). Discriminative reranking for natural language parsing. ICML 17 (pp.
175–182).
Collins, M. (2001). Parameter estimation for statistical parsing models: Theory and prac-
tice of distribution-free methods. IWPT.
Collins, M. (2004). Parameter estimation for statistical parsing models: Theory and prac-
tice of distribution-free methods. In H. Bunt, J. Carroll and G. Satta (Eds.), New devel-
opments in parsing technology. Kluwer.
Corduneanu, A., & Jaakkola, T. (2003). On information regularization. Proc. UAI.
Cormen, T., Leiserson, C., Rivest, R., & Stein, C. (2001). Introduction to algorithms. MIT
Press. 2nd edition.
Cover, T. M., & Thomas, J. A. (1991). Elements of information theory. New York: Wiley.
Cowell, R., Dawid, A., Lauritzen, S., & Spiegelhalter, D. (1999). Probabilistic networks
and expert systems. New York: Springer.
Crammer, K., & Singer, Y. (2001). On the algorithmic implementation of multiclass kernel-
based vector machines. Journal of Machine Learning Research, 2(5), 265–292.
Craven, M., DiPasquo, D., Freitag, D., McCallum, A., Mitchell, T., Nigam, K., & Slattery,
S. (1998). Learning to extract symbolic knowledge from the world wide web. Proc
AAAI98 (pp. 509–516).
Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines and
other kernel-based learning methods. Cambridge University Press.
DeGroot, M. H. (1970). Optimal statistical decisions. New York: McGraw-Hill.
Della Pietra, S., Della Pietra, V., & Lafferty, J. (1997). Inducing features of random fields.
IEEE Trans. on Pattern Analysis and Machine Intelligence, 19(4), 380–393.
Demaine, E. D., & Immorlica, N. (2003). Correlation clustering with partial information.
APPROX.
206 BIBLIOGRAPHY
Dempster, P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood from incomplete
data via the EM algorithm. Journal of the Royal Statistical Society, 39.
Devroye, L., Gy¨ orfi, L., & Lugosi, G. (1996). Probabilistic theory of pattern recognition.
New York: Springer-Verlag.
Duda, R. O., Hart, P. E., & Stork, D. G. (2000). Pattern classification. New York: Wiley
Interscience. 2nd edition.
Durbin, R., Eddy, S., Krogh, A., & Mitchison, G. (1998). Biological sequence analysis.
Cambridge University Press.
Edmonds, J. (1965). Maximum matching and a polyhedron with 0-1 vertices. Journal of
Research at the National Bureau of Standards, 69B, 125–130.
Egghe, L., & Rousseau, R. (1990). Introduction to informetrics. Elsevier.
Emanuel, D., & Fiat, A. (2003). Correlation clustering - minimizing disagreements on
arbitrary weighted graphs. ESA.
Fariselli, P., & Casadio, R. (2001). Prediction of disulfide connectivity in proteins. Bioin-
formatics, 17, 957–964.
Fariselli, P., Martelli, P., & Casadio, R. (2002). A neural network-based method for pre-
dicting the disulfide connectivity in proteins. Knowledge based intelligent information
engineering systems and allied technologies (KES 2002), 1, 464–468.
Fariselli, P., Ricobelli, P., & Casadio, R. (1999). Role of evolutionary information in pre-
dicting the disulfide-bonding state of cysteine in proteins. Proteins, 36, 340–346.
Fiser, A., & Simon, I. (2000). Predicting the oxidation state of cysteines by multiple se-
quence alignment. Bionformatics, 3, 251–256.
Forsyth, D. A., & Ponce, J. (2002). Computer vision: A modern approach. Prentice Hall.
Frasconi, P., Gori, M., & Sperduti, A. (1998). A general framework for adaptive structures.
IEEE Transactions on Neural Networks.
BIBLIOGRAPHY 207
Frasconi, P., Passerini, A., & Vullo, A. (2002). A two stage svm architecture for predicting
the disulfide bonding state of cysteines. Proceedings of IEEE Neural Network for signal
processing conference.
Freund, Y., & Schapire, R. (1998). Large margin classification using the perceptron algo-
rithm. Computational Learing Theory.
Friedman, N., Getoor, L., Koller, D., & Pfeffer, A. (1999). Learning probabilistic relational
models. Proc. IJCAI99 (pp. 1300–1309). Stockholm, Sweden.
Friess, T., Cristianini, N., & Campbell, C. (1998). The kernel adatron algorithm: a fast and
simple learning procedure for support vector machine. Proc. ICML.
Gabow, H. (1973). Implementation of algorithms for maximum matching on nonbipartite
graphs. Doctoral dissertation, Stanford University.
Garey, M. R., & Johnson, D. S. (1979). Computers and intractability. Freeman.
Geman, S., & Johnson, M. (2002). Dynamic programming for parsing and estimation of
stochastic unification-based grammars. Proceedings of the 40th Annual Meeting of the
Association for Computational Linguistics.
Getoor, L., Friedman, N., Koller, D., & Taskar, B. (2002). Learning probabilistic models
of link structure. Journal of Machine Learning Research, 8.
Getoor, L., Segal, E., Taskar, B., & Koller, D. (2001). Probabilistic models of text and link
structure for hypertext classification. Proc. IJCAI01 Workshop on Text Learning: Beyond
Supervision. Seattle, Wash.
Greig, D. M., Porteous, B. T., & Seheult, A. H. (1989). Exact maximum a posteriori
estimation for binar images. J. R. Statist. Soc. B, 51.
Guestrin, C., Koller, D., Parr, R., & Venkataraman, S. (2003). Efficient solution algorithms
for factored mdps. Journal of Artificial Intelligence Research, 19.
Gusfield, D. (1997). Algorithms on strings, trees, and sequences : Computer science and
computational biology. Cambridge University Press.
208 BIBLIOGRAPHY
Harrison, P., & Sternberg, M. (1994). Analysis and classification of disulphide connectivity
in proteins. the entropic effect of cross-linkage. Journal of Molecular Biology, 244.
Hastie, T., Tibshirani, R., & Friedman, J. (2001). The elements of statistical learning. New
York: Springer-Verlag.
Haussler, D. (1999). Convolution kernels on discrete structures (Technical Report). UC
Santa Cruz.
Heuberger, C. (2004). Inverse combinatorial optimization: Asurvey on problems, methods,
and results. Journal of Combinatorial Optimization, 8.
Hochbaum, D. S. (Ed.). (1997). Approximation algorithms for NP-hard problems. PWS
Publishing Company.
Joachims, T. (1999). Transductive inference for text classification using support vector
machines. Proc. ICML99 (pp. 200–209). Morgan Kaufmann Publishers, San Francisco,
US.
Johnson, M. (2001). Joint and conditional estimation of tagging and parsing models. ACL
39.
Johnson, M., Geman, S., Canon, S., Chi, Z., & Riezler, S. (1999). Estimators for stochastic
“unification-based” grammars. Proceedings of ACL 1999.
Kabsch, W., & Sander, C. (1983). Dictionary of protein secondary structure: Pattern recog-
nition of hydrogen-bonded and geometrical features.
Kaplan, R., Riezler, S., King, T., Maxwell, J., Vasserman, A., & Crouch, R. (2004). Speed
and accuracy in shallow and deep stochastic parsing. Proceedings of HLT-NAACL’04.
Kassel, R. (1995). A comparison of approaches to on-line handwritten character recogni-
tion. Doctoral dissertation, MIT Spoken Language Systems Group.
Kivinen, J., & Warmuth, M. (1997). Exponentiated gradient versus gradient descent for
linear predictors. Information and Computation, 132(1), 1–63.
BIBLIOGRAPHY 209
Klein, D., & Manning, C. D. (2003). Accurate unlexicalized parsing. ACL 41 (pp. 423–
430).
Kleinberg, J., & Tardos, E. (1999). Approximation algorithms for classification problems
with pairwise relationships: Metric labeling and Markov random fields. FOCS.
Kleinberg, J. M. (1999). Authoritative sources in a hyperlinked environment. Journal of
the ACM, 46(5), 604–632.
Klepeis, J., & Floudas, C. (2003). Prediction of β-sheet topology and disulfide bridges in
polypeptides. Journal of Computational Chemistry, 24, 191–208.
Koller, D., & Pfeffer, A. (1998). Probabilistic frame-based systems. Proc. AAAI98 (pp.
580–587). Madison, Wisc.
Kolmogorov, V., & Zabih, R. (2002). What energy functions can be minimized using graph
cuts? PAMI.
Kumar, S., & Hebert, M. (2003). Discriminative fields for modeling spatial dependencies
in natural images. NIPS.
Lafferty, J., McCallum, A., & Pereira, F. (2001). Conditional random fields: Probabilistic
models for segmenting and labeling sequence data. ICML.
Lafferty, J., Zhu, X., & Liu, Y. (2004). Kernel conditional random fields: Representation
and clique selection. ICML.
Lawler, E. (1976). Combinatorial optimization: Networks and matroids. New York: Holt,
Rinehart and Winston.
Leslie, C., Eskin, E., & Noble, W. (2002). The spectrum kernel: a string kernel for svm
protein classification. Proc. Pacific Symposium on Biocomputing.
Liu, Z., & Zhang, J. (2003). On inverse problems of optimum perfect matching. Journal
of Combinatorial Optimization, 7.
210 BIBLIOGRAPHY
Lodhi, H., Shawe-Taylor, J., Cristianini, N., & Watkins, C. (2000). Text classification using
string kernels. NIPS.
Luenberger, D. (1997). Investment science. Oxford University Press.
Manning, C., &Sch¨ utze, H. (1999). Foundations of statistical natural language processing.
Cambridge, Massachusetts: The MIT Press.
Marcus, M. P., Santorini, B., & Marcinkiewicz, M. A. (1993). Building a large annotated
corpus of English: The Penn Treebank. Computational Linguistics, 19(2), 313–330.
Markowitz, H. (1991). Portfolio selection: Efficient diversification of investments. Basil
Blackwell.
Martelli, P., Fariselli, P., Malaguti, L., & Casadio, R. (2002). Prediction of the disulfide-
bonding state of cysteines in proteins at 88% accuracy. Protein Science, 11, 27359.
Martin, D., Fowlkes, C., Tal, D., & Malik, J. (2001). A database of human segmented
natural images and its application to evaluating segmentation algorithms and measuring
ecological statistics. Proc. ICCV.
Matsumura, M., Signor, G., &Matthews, B. (1989). Substantial increase of protein stability
by multiple disulfide bonds. Nature, 342, 291:293.
Matusov, E., Zens, R., & Ney, H. (2004). Symmetric word alignments for statistical ma-
chine translation. Proc. COLING.
McCallum, A. (2003). Efficiently inducing features of conditional random fields. Proc.
UAI.
McCallum, A., & Wellner, B. (2003). Toward conditional models of identity uncertainty
with application to proper noun coreference. IJCAI Workshop on Information Integration
on the Web.
Mitchell, T. (1997). Machine learning. McGraw-Hill.
BIBLIOGRAPHY 211
Miyao, Y., & Tsujii, J. (2002). Maximum entropy estimation for feature forests. Proceed-
ings of Human Language Technology Conference (HLT 2002).
Murphy, K. P., Weiss, Y., & Jordan, M. I. (1999). Loopy belief propagation for approximate
inference: an empirical study. Proc. UAI99 (pp. 467–475).
Needleman, S., & Wunsch, C. (1970). A general method applicable to the search for
similarities in the amino acid sequences of two proteins. Journal of Molecular Biology,
48.
Nemhauser, G. L., & Wolsey, L. A. (1999). Integer and combinatorial optimization. New
York: J. Wiley.
Neville, J., & Jensen, D. (2000). Iterative classification in relational data. Proc. AAAI-2000
Workshop on Learning Statistical Models from Relational Data (pp. 13–20).
Ng, A., & Jordan, M. (2001). On discriminative vs. generative classifiers: A comparison
of logistic regression and naive Bayes. NIPS.
Ng, A. Y., & Russell, S. (2000). Algorithms for inverse reinforcement learning. Proc.
ICML.
Nigam, K., McCallum, A., Thrun, S., & Mitchell, T. (2000). Text classification from
labeled and unlabeled documents using em. Machine Learning, 39.
Nocedal, J., & Wright, S. J. (1999). Numerical optimization. New York: Springer.
Papadimitriou, C., & Steiglitz, K. (1982). Combinatorial optimization: Algorithms and
complexity. Englewood Cliffs, NJ: Prentice-Hall.
Pearl, J. (1988). Probabilistic reasoning in intelligent systems. San Francisco: Morgan
Kaufmann.
Pinto, D., McCallum, A., Wei, X., &Croft, W. B. (2003). Table extraction using conditional
random fields. Proc. ACM SIGIR.
212 BIBLIOGRAPHY
Platt, J. (1999). Using sparseness and analytic QP to speed training of support vector
machines. NIPS.
Potts, R. B. (1952). Some generalized order-disorder transformations. Proc. Cambridge
Phil. Soc., 48.
Quinlan, J. R. (2001). Induction of decision trees. Machine Learning, 1, 81–106.
Schrijver, A. (2003). Combinatorial optimization: Polyhedra and efficiency. Springer.
Sha, F., & Pereira, F. (2003). Shallow parsing with conditional random fields. Proc. HLT-
NAACL.
Shawe-Taylor, J., Bartlett, P. L., Williamson, R. C., & Anthony, M. (1998). Structural
risk minimization over data-dependent hierarchies. IEEE Trans. on Information Theory,
44(5), 1926–1940.
Shen, L., Sarkar, A., & Joshi, A. K. (2003). Using ltag based features in parse reranking.
Proc. EMNLP.
Slattery, S., & Mitchell, T. (2000). Discovering test set regularities in relational domains.
Proc. ICML00 (pp. 895–902).
Smith, T., & Waterman, M. (1981). Identification of common molecular subsequences. J.
Mol. Biol., 147, 195–197.
Sutton, C., Rohanimanesh, K., & McCallum, A. (2004). Dynamic conditional random
fields: Factorized probabilistic models for labeling and segmenting sequence data. Proc.
ICML.
Szummer, M., & Jaakkola, T. (2001). Partially labeled classification with Markov random
walks. Proc. NIPS.
Tarantola, A. (1987). Inverse problem theory: Methods for data fitting and model parame-
ter estimation. Amsterdam: Elsevier.
BIBLIOGRAPHY 213
Taskar, B., Abbeel, P., & Koller, D. (2002). Discriminative probabilistic models for rela-
tional data. UAI.
Taskar, B., Chatalbashev, V., & Koller, D. (2004a). Learning associative Markov networks.
Proc. ICML.
Taskar, B., Guestrin, C., & Koller, D. (2003a). Max margin Markov networks. NIPS.
Taskar, B., Klein, D., Collins, M., Koller, D., & Manning, C. (2004b). Max margin parsing.
EMNLP.
Taskar, B., Segal, E., & Koller, D. (2001). Probabilistic classification and clustering in
relational data. Proc. IJCAI01 (pp. 870–876). Seattle, Wash.
Taskar, B., Wong, M., Abbeel, P., & Koller, D. (2003b). Link prediction in relational data.
Proc. NIPS.
Toutanova, K., Klein, D., Manning, C. D., & Singer, Y. (2003). Feature-rich part-of-speech
tagging with a cyclic dependency network. NAACL 3 (pp. 252–259).
Tsochantaridis, I., Hofmann, T., Joachims, T., & Altun, Y. (2004). Support vector machine
learning for interdependent and structured output spaces. Twenty-first international con-
ference on Machine learning.
Valiant, L. G. (1979). The complexity of computing the permanent. Theoretical Computer
Science, 8, 189–201.
Vapnik, V. (1995). The nature of statistical learning theory. New York, New York:
Springer-Verlag.
Varma, A., & Palsson, B. (1994). Metabolic flux balancing: Basic concepts, scientific and
practical use. Bio/Technology, 12.
Vazquez, A., Flammini, A., Maritan, A., & Vespignani, A. (2003). Global protein function
prediction from protein-protein interaction networksh. Nature Biotechnology, 6.
214 BIBLIOGRAPHY
Veksler, O. (1999). Efficient graph-based energy minimization methods in computer vision.
Doctoral dissertation, Cornell University.
Vullo, A., & Frasconi, P. (2004). Disulfide connectivity prediction using recursive neural
networks and evolutionary information. Bioinformatics, 20, 653–659.
Wahba, G., Gu, C., Wang, Y., & Chappell, R. (1993). Soft classification, a.k.a. risk estima-
tion, via penalized log likelihood and smoothing spline analysis of variance. Computa-
tional Learning Theory and Natural Learning Systems.
Wainwright, M., Jaakkola, T., & Willsky, A. (2002). Map estimation via agreement on
(hyper)trees: Message-passing and linear programming approaches. Proc. Allerton Con-
ference on Communication, Control and Computing.
Wainwright, M., & Jordan, M. I. (2003). Variational inference in graphical models: The
view from the marginal polytope. Proc. Allerton Conference on Communication, Control
and Computing.
Weston, J., & Watkins, C. (1998). Multi-class support vector machines (Technical Re-
port CSD-TR-98-04). Department of Computer Science, Royal Holloway, University of
London.
Xing, E., Ng, A., Jordan, M., & Russell, S. (2002). Distance metric learning, with applica-
tion to clustering with side-information. NIPS.
Yang, Y., Slattery, S., & Ghani, R. (2002). A study of approaches to hypertext categoriza-
tion. Journal of Intelligent Information Systems, 18(2).
Yedidia, J., Freeman, W., & Weiss, Y. (2000). Generalized belief propagation. NIPS.
Younger, D. H. (1967). Recognition and parsing of context-free languages in time n
3
.
Information and Control, 10, 189–208.
Zhang, J., & Ma, Z. (1996). A network flow method for solving inverse combinatorial
optimization problems. Optimization, 37, 59–72.
BIBLIOGRAPHY 215
Zhang, T. (2002). Covering number bounds of certain regularized linear function classes.
Journal of Machine Learning Research, 2, 527–550.
Zhu, J., & Hastie, T. (2001). Kernel logistic regression and the import vector machine.
Proc. NIPS.
Zhu, X., Ghahramani, Z., & Lafferty, J. (2003). Semi-supervised learning using gaussian
fields and harmonic functions. Proc. ICML.

c Copyright by Ben Taskar 2005 All Rights Reserved

ii

I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy.

Daphne Koller
Computer Science Department Stanford University

(Principal Advisor)

I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy.

Andrew Y. Ng
Computer Science Department Stanford University

I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy.

Fernando Pereira
Computer Science Department University of Pennsylvania

Approved for the University Committee on Graduate Studies:

iii

iv

Anat. Mark and Tsilya.To my parents. . and my love.

.

Such complex answers may consist of multiple interrelated decisions that must be weighed against each other to arrive at a globally satisfactory and consistent solution to the question. grammatical. The acquired expertise must now be used to address tasks that demand answers as complex as the questions. and track motion of articulated bodies. reconstruct 3D shapes from stereo and video. spatial constraints and correlations. like standardized testing. or translation into another language. we analyze genetic sequences to predict 3D structure of proteins. In computational biology. has placed the heaviest emphasis on complex questions with simple answers. coherent analysis of a sentence. combinatorial optimization problems such as weighted matchings and graph-cuts. We typically handle the exponential explosion of possible answers by building models that compactly capture the structural properties of the problem: sequential. chemical. find global alignment of related DNA strings.Abstract Most questions require more than just true-false or multiple-choice answers. recursive language models such as context free grammars. This thesis presents a discriminative estimation framework for structured models based on the large margin principle underlying support vector machines. In computer vision. Yet supervised learning. temporal. In natural language processing. we segment complex objects in cluttered scenes. Our framework defines a suite of efficient learning algorithms that rely on the expressive power of convex optimization to compactly capture inference or vii . parse tree. Intuitively. and recognize functional portions of a genome. Such structured models include graphical models such as Markov networks (Markov random fields). the large-margin criterion provides an alternative to probabilistic. such as its corresponding part-of-speech sequence. likelihood-based estimation methods by concentrating directly on the robustness of the decision boundary of a model. we often need to construct a global.

Using an efficient online-style algorithm that exploits inference in the model and analytic updates. which compactly define templates for Markov networks for domains with relational structure: objects. convex formulation for large-margin estimation of Markov networks with sequence and other low-treewidth structure. natural language parsing.solution optimality in structured models. We analyze the theoretical generalization properties of max-margin estimation in Markov networks and derive a novel type of bound on structured error. In addition to graphical models. For some of these models. disulfide connectivity prediction in proteins. including handwriting recognition. While likelihood-based methods are believed to be intractable for AMNs over binary variables. We use combinatorial properties of weighted matchings to develop an exact. which are undirected probabilistic graphical models widely used to efficiently represent and reason about joint multivariate distributions. attributes. We define an important subclass of Markov networks. our framework applies to a wide range of other models: We exploit context free grammar structure to derive a compact max-margin formulation that allows high-accuracy parsing in cubic time by using novel kinds of lexical information. tuning one to the other in a joint optimization. Seamless integration of kernels with graphical models allows efficient. efficient formulation for learning to match and apply it to prediction of disulfide connectivity in proteins. we solve very large estimation problems. email organization and image segmentation. These empirical evaluations show significant improvements over state-of-the-art methods and promise wide practical use for our framework. which captures positive correlations between variables and permits exact inference which scales up to tens of millions of nodes and edges. We use graph decomposition to derive an exact. accurate prediction in real-world tasks. alternative estimation methods are intractable. compact. Finally. viii . which tightly integrates metric learning with the clustering algorithm. The largest portion of the thesis is devoted to Markov networks. we derive a max-margin formulation for learning the scoring metric for clustering from clustered training data. 3D terrain classification. hypertext categorization. relations. our framework allows exact estimation of such networks of arbitrary connectivity and topology. We also introduce relational Markov networks (RMNs). We describe experimental applications to a diverse range of tasks. associative Markov networks (AMNs).

Xavier Boyen. Her tireless pursuit of excellence in research. David Vickrey. Ron Parr. teaching. I have learned a great deal from their work and their influence on this thesis is immense. Stephen’s book on convex optimization is the primary source of many of the insights and derivations in this thesis. as well as my defense committee members. I would like to thank my thesis committee.Acknowledgements I am profoundly grateful to my advisor. Daphne Koller. Uri Nodelman. Sebastian introduced me to the fascinating problems in 3D vision and robotics. In particular. Christian Shelton. Luke Biewald. I am very lucky to have been a part of it and shared the company. Finally. the ideas. pushing ideas to their full development. Uri Lerner. advising. Carlos Guestrin. Andrew’s research on clustering and classification has informed many of the practical problems addressed in the thesis. I ix . Lise Getoor. I am indebted to Daphne for priceless and copious advice about selecting interesting problems. Haidong Wang and Ming Fai Wong. Simon Tong. When I started my research in machine learning with Daphne. Dirk Ormoneit. Fernando’s work on Markov networks has inspired my focus on this subject. Daphne’s research group is an institution in itself. Jean-Claude Latombe was my first research advisor when I was still an undergraduate at Stanford. Andrew Ng and Fernando Pereira. Stephen Boyd and Sebastian Thrun. Drago Anguelov. Alexis Battle. for their excellent suggestions and thought-provoking questions. Vassil Chatalbashev. the questions and the expertise of Pieter Abbeel. His thoughtful guidance and quick wit were a great welcome and inspiration to continue my studies. making progress on difficult ones. Eran Segal. and every other aspect of her academic work is truly inspirational. Gal Chechik. writing and presenting results in an engaging manner.

Vassil Chatalbashev. always late. David Vickrey and Ming Fai Wong. enlightened much of my work. languages and generalization bounds. Dan has an incredible gift of making ideas work. I was lucky to meet Michael Collins. good cheer and the great knack for always coming through. Lise Getoor. for his company and ideas. Vasco Chatalbashev. anywhere: gym. Audrey Gasch. Chris Manning. while I was a summer intern at AT&T. avid grillmaster x . Ogromnoye spasibo to my officemate. to mention a few. tennis court. has been a constant source of encouragement and inspiration. for numerous enlightening discussions that we have had anytime. fellow photography fanatic. Blagodarya vi mnogo for another great Bulgarian I have the good luck to know. Merci beaucoup to Pieter Abbeel. Peter Bartlett. David McAllester. Drago Anguelov. Dan Klein. I am very glad I had the chance to work with him on the max-margin parsing project. Luke Biewald. David’s wideranging research on probability and logic. Michael Littman and David McAllester all in one place. Muchas gracias to Carlos Guestrin. biking to EV. Collins’ work on learning large-margin tagging and parsing models motivated me to look at such methods for other structured models. Geremy Heitz. hard work and dedication: Pieter Abbeel. Nir Friedman. Chris’ healthy scepticism and emphasis on the practical has kept me from going down many dead-end paths. ever-ready to checkmate. Drago Anguelov. with his endearing combination of humor and warmth. explaining why other ideas do not work. I could not wish for more caring advice and encouragement than she has given me all these years. Lise has been a great friend and my other big sister throughout my graduate life. my personal climbing instructor. Carlos Guestrin. sportsmate. lightness and depth.had the privilege and pleasure to work with Nir Friedman. Daphne Koller. my sounding board for many ideas in convex optimization and learning. who is always willing to share with me everything from research problems to music and backpacking treks to Russian books and Belgian beer. Lise Getoor and Eran Segal. and better yet. Michael Littman. They have taught me to remain steadfast when elegant theories meet the ragged edge of real problems. Eran Segal. (Almost) everything I know about natural language processing I learned from Chris Manning and Dan Klein. Michael Collins. over cheap Thai food. I would like to express my appreciation and gratitude to all my collaborators for their excellent ideas.

for opening his ear and heart to me when it matters most. I am grateful for bringing me so much joy and love. kind heart and excellent work.and exploiter-of-structure extraordinaire. tears and laughter. Mark and Tsilya. My parents. To my tireless partner in work and play. xi . thank you for making me happy next to you. I will not list all of you here. have given me unbending support and constant encouragement. but worked hard to keep my relative sanity throughout. my brother-in-law. and my sweet niece. Talia. for his sense of humor. but my gratitude to you is immense. hardship and love. Anat Caspi. Many thanks to my friends who have had nothing to do with work in this thesis. Ilya. I thank them for all the sacrifices they have made to ensure that their children would have freedom and opportunities they never had in Soviet Union. Mh-goi-saai to Ming Fai Wong. To my sister. Anna.

xii .

20 Logistic dual and maximum entropy . . . . . . .3 2. . . . . . . . . . . . .4 2. . . . . Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 1. . . . . . . . . . . .6 Supervised learning . . . . . . . . .5 15 16 Classification with generalized linear models . . . . . . . .1 2. . . . . . . .2 2. . . . . . . . . 25 xiii .4 1. . .3 1. . . . . . . . .5 1. . . . . . . . . . . . . . . . Structured models . . . . . . . . . . . . . . . . . . . .Contents Abstract Acknowledgements 1 Introduction 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 1. . . . . . . . . 11 Previously published work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Support vector machines . . 13 I Models and methods 2 Supervised learning 2. .1 Probabilistic models: generative and conditional . . . . . . . . 21 SVM dual and kernels . . . . . . . . . . . . . . Complex prediction problems . . . . . . . . . . . . . . . . . . . . . vii ix 1 2 4 6 8 Thesis outline . . . . . . . . . . . . . . . . . 22 24 3 Structured models 3. . . . 19 Logistic regression . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .1. . . . . . . . . . .1 4. . . . . . . . . . . . . . .1. . . . . . . . . . . . 27 Markov networks . . . . . . . . . . . . .3 Prediction models: normalized and unnormalized . . . .3 4. . . . . . . 34 3. . . . . . . . . . 73 Conclusion . . . . . 28 Inference . . . . . . . . . . . . . . . .4 Max-margin estimation . 37 Combinatorial problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 II Markov networks 57 58 5 Markov networks 5. . . 48 Constraint generation . . 69 Generalization bound . . . . . . . 52 Approximations: upper and lower bounds . . . . . . . . . . . . . . . . . . . .1 4. . . . . 43 Min-max formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . 61 M3 N dual and kernels . . . . . . . . . . 70 Related work . . . . . . . 65 Untriangulated models . 31 Linear programming MAP inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 Related work . .3. . . . . . . . . . . . . . . . . .2. . . . . . . . . . . . . . . . .1 4. . . . . . . . . . . . . . . . . . . 50 Constraint strengthening . . . . . . . . . . . . 59 Maximum margin estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 3. . . . . . . . . . . . . . . .3 5. . . .2 5. . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . . . . . 74 75 6 M3 N algorithms and experiments 6. . . . . . . . . . .5 Context free grammars . . . .1 Solving the M3 N QP . .2 4. . . . . . . . . . . . . . . . . . .2 4. . . . . . . . . . . . . . . . . . . . . . .2 4. . . . . . . . . . .3. . 40 42 4 Structured maximum margin estimation 4. . . . . . . . . . . . . . . . . . . . 43 Certificate formulation . . . .6 5. . . . . . . . . . . . 75 SMO . . . . . . . .5 5. .1 3. . . . . . . 78 xiv . . . . . . . . . .1 5. . . . . .3 Representation . . . . . . .4 5. . . .3. . . . . . . . . . . . . . . . . . . . . . . 27 3. . . . . . . . . . .1 6. . . . . . . 53 Conclusion . . . . . . . .2 3. . . . . . . . . . .7 Maximum likelihood estimation . . . .2 3. .

. . . . . . . . . . . 90 LP Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 6. .6 Related work . . . .3 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . .7 Max-margin estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 7. . . . . . . . . . . . . . . . . . . . . . 97 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 8. . . . . . . . . . clustering 9 Context free grammars 9. . . . . . . . . . . . . . . . . . . .2 8. . . . . . . . . . . . . .1 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 Relational Markov networks . . . . . . . . 121 8. . . . . . . . . . . . .5 8. . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . .3 6. . . . 87 Conclusion . . . . . .3 Associative networks . . . . . . . . . . . . . . . . . . . . . .1 8. . . . . . . . . 104 Conclusion . . . . . . . . . . . . . 124 III Broader applications: parsing. . . . . . . . . . . . . . . . . . . . 93 7. . . . . . . . . . . 128 xv . . . 81 Experiments . . 123 Conclusion . .4. . . . . . . . . . . . . . . . . . . . . .4 7. . . . . . 95 7. . . . . . . . . . . . . 119 Link model . . . .2 6. . . . . . . . . . . . . 104 109 8 Relational Markov networks 8. . . . . . . . .1. .1 7. . . . . . . 91 Min-cut inference . .1 8. . . 79 Structured SMO . . . matching. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 7. .6 7. . . . . . . . . . . . . . .4 Relational classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 Multi-class case . . . .3. . 120 Cocite model . . . . . . . . . . . . . . . . . . . . . 116 Experiments .1 127 128 Context free grammar model . . . . . . . .3 Flat models . . .4. . . . . . .3 6. . . . . . . . . . . . . .2 Graph construction . . . . . . . . . .6. . . . . . . . . . . . . . . . . . . 85 Related work . . . . . . . . . . . . . . . 113 Approximate inference and learning . . . . . . . . . . . . . . .4 Selecting SMO pairs .2 8. . . . . . . . .3. . . . . . 87 89 7 Associative Markov networks 7. . . . . 99 Related work . . . . . . . . . .

. . . . . . . . . . . . . .1 9. .9. . . . . . . . 168 11. . . . . . . . . . . . 167 11. . . . . .2 Learning to match . . . . . . . . . . . . . . . . . . . . . . . . . 145 10. . . . . . . . . . .4 Certificate formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 133 9. . . . . . .2 Email clustering . . . . . . . . . . .8 Conclusion . . . . . . . . . . . 164 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 11. . . . . .4. . . . . . . . . . . . . . . . . . . . . . . . 150 10. . 163 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 9. . . . .2 Learning formulation . . . . . . . . . . . . . . . . . . . .7 Related work . .2 Semidefinite programming relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Related work . 172 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 Maximum margin estimation . .1 Clustering formulation . . . . . . . .2 9. 138 Experiments . . . . . . . . .3 Min-max formulation . .3. . . . . . . . 142 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 11. . . . . . . . . 159 11 Correlation clustering 160 11. . . . . . .4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Kernels .4. . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . . . . . . . . . . . . . . . 173 xvi . . . . 143 144 10 Matchings 10. 162 11. . . .3 Context free parsing .6 Experiments . . . . . . . . . . . . . . . . . . .6 Conclusion . . . . . . . . . . . . . . 132 Discriminative parsing models . . . . . .1. . . . . . . . . . . 154 10. .7 Structured SMO for CFGs . . . . . . . . . . . . 146 10. . . . . . . . . . . . . . . . . . . . . . . . . . .3 Dual formulation and kernels . . 168 11.4 9. . . . . . . . . . 153 10. . . . . . . . 157 10. . . . .1 Linear programming relaxation . . . . . . . . . . . . . .3. . .3 Image segmentation . . . . . . . . . 138 Related work . . . . . . . . . . . . . . . . . . .1 Irrelevant features . . . . .1 Disulfide connectivity prediction . . .1. . . . . . . . . . . . . .5 9. . . . . . . . . . . 148 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 9. . . . . . . . . . . . . . . . . . . . . . . . 168 11. . . . . .

. . . . relational . . . . .1 Binary AMNs . . . . . . . . . .IV Conclusions and future directions 12 Conclusions and future directions 175 176 12. . . . . . . . . 182 12. . . . . . . . . . . . . . . . . . . . . . . . . 177 12. . . .2. . . . . 197 A. . . 183 12. . . . . . 195 A.5.1 Structured maximum margin estimation . . .1.1 Binary classification . . . . . .1 Summary of contributions . . . . . . . .2 AMN proofs and derivations .2 Multi-class AMNs . . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . 183 12. . . . .2. .2 Markov networks: max-margin. . . . 185 12. . . . . . . . clustering .2. . . . . . . . . . . . . . . 184 12. . .3 Broader applications: parsing. . . . . .1 Theoretical analysis and optimization algorithms . . . . . . . associative. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 A. . . .1. . . . . 187 A Proofs and derivations 188 A. . . . . . . . . . . . . . . . 176 12.1 Proof of Theorem 5. . . . . . . . . 199 Bibliography 202 xvii . 180 12. . . . . . . . . . . . .2 Structured classification . . . . . . . . . . 195 A. . . . . . . . . . . . . . . . . . . . . . 188 A. . .1. . .2. . . .1 .1. . . . . . . . .3 More general learning settings . . . . matching. . .3 Future . . . .2 Novel prediction tasks . . . . . . . . .3 Derivation of the factored primal and dual max-margin QP . . . . . . . . . . . .1. . . . . 190 A. . . . . .2. .2 Extensions and open problems . . . . . . . . . . . . .

. . . . 140 10. . . . . . . . . . . . . . . .1 Bond connectivity prediction results . . . .1 9. . . . . . . . . . . . . . . . .List of Tables 9. . . . . . . . . .2 Parsing results on development set . . . . . . . . . 156 xviii . . . . . . . . . . . . 140 Parsing results on test set . .

. . . . . . . . . . . 86 xix .6 4. . . . . . . . . . 20 Handwritten word recognition .4 6. 32 Diamond network junction tree . . . .3 6. . . . . . . . . . . .2 3. . . 64 Diamond Markov network . . . . .2 2. . . . . . . . . . . . . . . . . . . .1 2. . . . . . . . . . . . . . . . . . . . 77 SMO subproblem . . . . . . . . 3 4 Handwritten character recognition .5 6. . . . . . . . . . . . . . . . . 35 Example parse tree . . . . . . . . . . . . . . . . .List of Figures 1. . . . . .1 6. . . . . . . . . . . . . . . . 82 Structured SMO pair selection . . . . . . . . . . . . . . . . . . . . . . . . 51 A constraint generation algorithm. . . . . . . . . . . . . . . . . .1 3. . . . .6 Supervised learning setting . . . . . .2 5. . . . . . . . . . . . . . . .1 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 Exact and approximate constraints for max-margin estimation . . . . . . . . . . . . . . . . . Examples of complex prediction problems . .2 6. . . . . . . . . . . . . .3 3. . . .2 3. . . . . . . . . 29 Diamond Markov network . . . . . . . . . . . . . . . . . . . . 85 OCR results . . . . . . . . . . . . . . . . 52 Chain M3 N example .2 6. . . . . 25 Chain Markov network . . . . . .1 4. . . . .5 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 3. . . . . 19 Classification loss and upper bounds . . . . . . . . . . . . . 34 Marginal agreement . . . . . . . . . . . . . . 80 Structured SMO diagram . . . . . . 69 Block-coordinate ascent . . . . . . . . . . . . . . . . . . . . 79 SMO pair selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 1. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .1 Learning to cluster with irrelevant features . . . . 120 WebKB results: link model . . . . . 147 10. .5 7. 103 Comparison of terrain classification models (zoomed-out) . . . . . . . .2 7. . . .9 8. . . . . . . . . . . . . . . . . . . . . . .1 PDB protein 1ANS . . . . . 169 11. . . . . . . . . . . . . . . . . . .8 7. . . . .2 Number of constraints vs. 171 11. . . . . . . . . .4 9. . . . . . . . . . . . . . . . . . . . . . . . . .6 7. . . . . . . . . . 96 Segbot: roving robot equipped with SICK2 laser sensors. . . . . . . . . . 106 Labeled portion of the test terrain dataset (Ground truth and SVM predictions)107 Labeled portion of the test terrain dataset (Voted-SVM and AMN predictions)108 Link model for document classification . . . . . . . . . . . . . . .2 8. . . . . . . . . . . . . . . 122 Two representations of a binary parse tree . . . . . . . . 102 Min-cut inference running times . . . . . 150 10. . . . .4 7. . . . .1 Min-cut graph construction . . . . . . . . .1 8. . . . . . . . . . . . .3 Learning curve for bond connectivity prediction . number of bonds . . . . . 100 3D laser scan range map of the Stanford Quad. . . . . . . . . . . 94 α-expansion algorithm . . . . . . . . . 157 11. . . . . . . . . . . . . . . . .3 8. .7 7. . . . . . . . . . . . . . .7. . . . . . . .3 7. . . . . .3 Two segmentations by different users . . . . . . 172 xx . . . . . . . . . . . . . . . . . . . . . . 101 Comparison of terrain classification models (detail) . . . . . . 130 10. . 170 11. . . . . . . . . . . 121 WebKB results: cocite model . . . 113 WebKB results: flat models . . . .2 Learning to organize email . . . . . . . . . . . . . . . . . .4 Two segmentations by different models . . . . . . . . . . . . . . . . . . . . . . .1 7. . . . . . . . . . . . . . . . . .

are performed almost effortlessly by a person. DNA sequence analysis. handwriting recognition. we encounter the need for very sophisticated decision making from our programs. Of course. As we work our way down that list. as opposed to. frequent grammatical constructions of the phrases.Chapter 1 Introduction The breadth of tasks addressed by machine learning is rapidly expanding. the individual letters. Machine learning offers an alternative to encoding all the intricate details of such a model from scratch. game playing. common ways to combine words into phrases. image analysis. financial analysis. but remain difficult and error-prone for computers. fraud detection. Computational models for handwriting recognition aim to capture the salient properties of the problem: typical shapes of the letters. the list of things we would like a computer to learn to do is much. etc. neuroscientists. One of its primary goals 1 . robot locomotion and many more. The complex synthesis of many levels of signal processing a person executes when confronted by a line of handwritten text is daunting. Major applications include medical diagnosis. say. scientific discovery. are primarily concerned with achieving acceptable speed and accuracy of recognition rather than modeling this complicated process with any biological verity. The reconstruction of an entire sentence from the photons hitting the retina off of each tiny patch of an image undoubtedly requires an elaborate interplay of recognition and representation of the pen-strokes. speech and handwriting recognition. whole words and constituent phrases. likely letter combinations that make up words. Computer scientists. for example. Some tasks. much longer.

2

CHAPTER 1. INTRODUCTION

is to devise efficient algorithms for training computers to automatically acquire effective and accurate models from experience. In this thesis, we present a discriminative learning framework and a novel family of efficient models and algorithms for complex recognition tasks in several disciplines, including natural language processing, computer vision and computational biology. We develop theoretical foundations for our approach and show a wide range of experimental applications, including handwriting recognition, 3-dimensional terrain classification, disulfide connectivity in protein structure prediction, hypertext categorization, natural language parsing, email organization and image segmentation.

1.1 Supervised learning
The most basic supervised learning task is classification. Suppose we wish to learn to recognize a handwritten character from a scanned image. This is a classification task, because we must assigns a class (an English letter from ‘a’ through ‘z’) to an observation of an object (an image). Essentially, a classifier is a function that maps an input (an image) to an output (a letter). In the supervised learning setting, we construct a classifier by observing labeled training examples, in our case, sample images paired with appropriate letters. The main problem addressed by supervised learning is generalization. The learning program is allowed to observe only a small sample of labeled images to produce an accurate classifier on unseen images of letters. More formally, let x denote an input. For example, a black-and-white image x can be represented as a vector of pixel intensities. We use X to denote the space of all possible inputs. Let y denote the output, and Y be the discrete space of possible outcomes (e.g., 26 letters ‘a’-‘z’). A classifier (or hypothesis) h is a function from X to Y, h : X → Y. We denote the set of all classifiers that our learning program can produce as H (hypothesis class). Then given a set of labeled examples {x(i) , y (i) }, i = 1, . . . , m, a learning program seeks to produce a classifier h ∈ H that will work well on unseen examples x, usually by finding h that accurately classifies training data. The diagram in Fig. 1.1 summarizes the supervised learning setting.

1.1. SUPERVISED LEARNING

3

Hypotheses

Labeled data

Learning Prediction New data

Figure 1.1: Supervised learning setting

The problem of classification has a long history and highly developed theory and practice (see for example, Mitchell [1997]; Vapnik [1995]; Duda et al. [2000]; Hastie et al. [2001]). The two most important dimensions of variation of classification algorithms is the hypothesis class H and the criterion for selection of a hypothesis h from H given the training data. In this thesis, we build upon the generalized linear model family, which underlies standard classifiers such as logistic regression and support vector machines. Through the use of kernels to implicitly define high-dimensional and even infinite-dimensional input representations, generalized linear models can approximate arbitrarily complex decision boundaries. The task of selecting a hypothesis h reduces to estimating model parameters. Broadly speaking, probabilistic estimation methods associate a joint distribution p(x, y) or conditional distribution p(y | x) with h and select a model based on the likelihood of the data [Hastie et al., 2001]. Joint distribution models are often called generative, while conditional models are called discriminative. Large margin methods, by contrast, select a model based on a more direct measure of confidence of its predictions on the training data called the margin [Vapnik, 1995]. The difference between these two methods is one of the key

4

CHAPTER 1. INTRODUCTION

The screen was a sea of red

RSCCPCYWGGCPW GQNCYPEGCSGPKV

brace

(a)

(b)

(c)

(d)

Figure 1.2: Examples of complex prediction problems (inputs-top, outputs-bottom):
(a) handwriting recognition [image → word]; (b) natural language parsing [sentence → parse tree]; (c) disulfide bond prediction in proteins [amino-acid sequence → bond structure (shown in yellow)]; (d) terrain segmentation [3D image → segmented objects (trees, bushes, buildings, ground)]

themes in this thesis. Most of the research has focused on the analysis and classification algorithms for the case of binary outcomes |Y| = 2, or a small number of classes. In this work, we focus on prediction tasks that involve not a single decision with a small set of outcomes, but a complex, interrelated collection of decisions.

1.2 Complex prediction problems
Consider once more the problem of character recognition. In fact, a more natural and useful task is recognizing words and entire sentences. Fig. 1.2(a) shows an example handwritten word “brace.” Distinguishing between the second letter and fourth letter (‘r’ and ‘c’) in isolation is actually far from trivial, but in the context of the surrounding letters that together form a word, this task is much less error-prone for humans and should be for computers as well. It is also more complicated, as different decisions must be weighed against each other to arrive at the globally satisfactory prediction. The space of all possible outcomes

1.2. COMPLEX PREDICTION PROBLEMS

5

Y is immense, usually exponential in the number of individual decisions, for example, the number of 5 letter sequences (265 ). However, most of these outcomes are unlikely given the observed input. By capturing the most salient structure of the problem, for example the strong local correlations between consecutive letters, we will construct compact models that efficiently deal with this complexity. Below we list several examples from different fields. • Natural language processing Vast amounts of electronically available text have spurred a tremendous amount of research into automatic analysis and processing of natural language. We mention some of the lower-level tasks that have received a lot of recent attention [Charniak, 1993; Manning & Sch¨ tze, 1999]. Part-of-speech tagging involves assigning each u word in a sentence a part-of-speech tag, such as noun, verb, pronoun, etc. As with handwriting recognition, capturing sequential structure of correlations between consecutive tags is key. In parsing, the goal is to recognize the recursive phrase structure of a sentence, such as verbal, noun and prepositional phrases and their nesting in relation to each other. Fig. 1.2(b) shows a parse tree corresponding to the sentence: “The screen was a sea of red” (more on this in Ch. 9). Many other problems, such as named-entity and relation extraction, text summarization, translation, involve complex global decision making. • Computational biology The last two decades have yielded a wealth of high-throughput experimental data, including complete sequencing of many genomes, precise measurements of protein 3D structure, genome-wide assays of mRNA levels and protein-protein interactions. Major research has been devoted to gene-finding, alignment of sequences, protein structure prediction, molecular pathway discovery [Gusfield, 1997; Durbin et al., 1998]. Fig. 1.2(c) shows disulfide bond structure (shown in yellow) we would like to predict from the amino-acid sequence of the protein (more on this in Ch. 10). • Computer vision As digital cameras and optical scanners become commonplace accessories, medical imaging technology produces detailed physiological measurements, laser scanners

Such models are encoded by a graph.2(d) shows a 3D laser range data image of the Stanford campus collected by a roving robot which we would like to segment into objects such as trees. parameterized scoring function. satellites and telescopes bring pictures of Earth and distant stars. and whose edges represent and quantify direct dependencies between them. Inference in these models refers to computing the highest scoring output given the input and usually involves dynamic programming or combinatorial optimization. 2002]. spatial. 7).k. Example tasks include object detection and segmentation. Markov networks compactly represent complex joint distributions of the label variables by modeling their local interactions. 1. 1. We focus on the models that compactly capture correlation and constraint structure inherent to many tasks. etc. the treewidth of the graph. The graphical structure of the models encodes the qualitative aspects of the distribution: direct dependencies as well as conditional independencies. • Context free grammars . Markov random fields) are extensively used to model complex sequential. typically through a compact.a. • Markov networks Markov networks (a.3 Structured models This wide range of problems have been tackled using various models and methods. and much more [Forsyth & Ponce. whose nodes represent the different object labels. Fig. a model assigns a score (or likelihood in probabilistic models) to each possible input/output pair (x. INTRODUCTION capture 3D environments. The quantitative aspect of the model is defined by the potentials that are associated with nodes and cliques of the graph. y). These problems involve labeling a set of related objects that exhibit local consistency. ground. we are flooded with images we would like our computer to analyze. (more on this in Ch. which we formally define in Ch. 3D reconstruction from stereo and video.6 CHAPTER 1. motion tracking. 3) is critical to efficient inference and learning in the model. bushes. buildings. and relational interactions in prediction problems arising in many fields. The graphical structure of the network (more precisely. Abstractly.

• Combinatorial structures Many important computational tasks are formulated as combinatorial optimization problems such as the maximum weight bipartite and perfect matching. . edge-cover. 1982. bipartite matchings. the non-terminal symbols (labels of internal nodes) correspond to syntactic categories such as noun phrase (NP). probabilistic CFGs construct a distribution over parse trees and sentences. The standard methods of estimation for Markov networks and context free grammars are based on maximum likelihood. The other combinatorial structures we consider and apply in this thesis include graph cuts and partitions. 1. For example. The productions define the set of syntactically allowed phrase structures (derivations).2(c) can be modeled by maximum weight perfect matchings. determiners (DT) and prepositions (IN). STRUCTURED MODELS 7 Context-free grammars are one of the primary formalisms for capturing the recursive structure of syntactic constructions [Manning & Sch¨ tze. 1. verbal phrase (VP) or prepositional phrase (PP) and part-of-speech tags like nouns (NN). For example. Cormen et al. verbs (VBD). to include any scheme that assigns scores to the output space Y and has a procedure for finding the optimal scoring y. both joint and conditional. and the prediction task reduces to finding the most likely tree given the sentence. u in Fig. V P → V P P P . A CFG consists of recursive productions (e. 1976. However. graph-cut.2.g. and many others [Lawler. where the weights define potential bond strength based on the local amino-acid sequence properties.1. maximum likelihood estimation of scoring function parameters for combinatorial structures is often intractable because of the problem of defining a normalized distribution over an exponential set of combinatorial structures. The context free restriction allows efficient inference and learning in such models. DT → The) that can be applied to derive a sentence of the language. By compactly defining a probability distribution over individual productions. Although the term ‘model’ is often reserved for probabilistic models. 2001]. Papadimitriou & Steiglitz. The terminal symbols (leaves) are the words of the sentence. we use the term model very broadly. 1999]. spanning tree.3. the disulfide connectivity prediction in Fig. and spanning trees..

The first formulation relies on the ability to express inference in the model as a compact convex optimization problem. INTRODUCTION 1. The seamless integration of kernels with graphical models allows us to create . ◦ Learning framework for structured models We propose a general framework for efficient estimation of models for structured prediction. We focus on those models where exact inference is tractable. These two formulations form the foundation which the rest of the thesis develops. compact. Low-treewidth Markov networks We use graph decomposition to derive an exact. vision and biology. we find parameters such that inference in the model (dynamic programming. Intuitively. ◦ Markov networks The largest portion of the thesis is devoted to novel estimation algorithms. taking advantage of prediction task structure. The second one only requires compactly expressing optimality of a given assignment according to the model and applies to a broader range of combinatorial problems. this framework builds upon the large margin estimation principle. a model class of choice in many structured prediction tasks in language. We develop general conditions under which exact large margin estimation is tractable and present two formulations for structured max-margin estimation that define compact convex optimization problems. generalization analysis and experimental validation for Markov networks.4 Contributions This thesis addresses the problem of efficient learning of high-accuracy models for complex prediction problems. convex learning formulation for Markov networks with sequence and other low-treewidth structure. We consider a very large class of structured models.8 CHAPTER 1. from Markov networks to context free grammars to combinatorial graph structures such as matchings and cuts. An alternative to likelihood-based methods. or can be efficiently approximated. representational extensions. combinatorial optimization) predicts the correct answers on the training data with maximum confidence.

for which likelihood methods are believed to be intractable. proportion of mislabeled pixels in an image) and uses a proof that exploits the graphical model structure. scale up to tens of millions of nodes and edges. For the non-binary case. our framework allows exact estimation of networks of arbitrary connectivity and topology. This bound is the first to address structured error (e. CONTRIBUTIONS 9 very rich models that leverage the immense amount of research in kernel design and graphical model decompositions for efficient. Generalization analysis We analyze the theoretical generalization properties of max-margin estimation in Markov networks and derive a novel margin-based bound for structured prediction.4. . Learning associative Markov networks (AMNs) We define an important subclass of Markov networks that captures positive correlations present in many domains. accurate prediction in real-world tasks. at run-time.1. Our online-style algorithm uses inference in the model and analytic updates to solve extremely large estimation problems. Scalable online algorithm We present an efficient algorithm for solving the estimation problem called Structured SMO.g. We also use approximate graph decomposition to derive a compact approximate formulation for Markov networks in which inference is intractable. we provide an approximation that works well in practice. Representation and learning of relational Markov networks We introduce relational Markov networks (RMNs). We show that for AMNs over binary variables. By constraining the class of Markov networks to AMNs. our models are learned efficiently and. We present an AMN-based method for object segmentation from 3D range data. which compactly define templates for Markov networks for domains with relational structure objects.

Experiments on synthetic and real-world data show the ability of the algorithm to learn an appropriate . We apply this class of models to classification of hypertext using hyperlink structure to define relations between webpages. The algorithm we propose uses kernels. ◦ Broader applications: parsing. we describe a formulation that tightly integrates metric learning with the clustering algorithm. perfect matchings for disulfide connectivity in protein structure prediction. Learning to cluster We derive a max-margin formulation for learning the affinity metric for clustering from clustered training data. in which exact inference is intractable. INTRODUCTION attributes. The graphical structure of an RMN is based on the relational structure of the domain. In contrast to algorithms that learn a metric independently of the algorithm that will be used to cluster the data. clustering The other large portion the thesis addresses a range of prediction tasks with very diverse models: context free grammars for natural language parsing. to derive an approximate learning formulation. matching. tuning one to the other in a joint optimization. graph partitions for clustering documents and segmenting images. relations. which makes it possible to efficiently embed the features in very high-dimensional spaces and achieve state-of-the-art accuracy. efficient algorithm for learning to match. Learning to parse We exploit context free grammar structure to derive a compact max-margin formulation and show high-accuracy parsing in cubic time by exploiting novel kinds of lexical information. We use approximate inference in these complex models. and can easily model complex interaction patterns over related entities. We show experimental evidence of the model’s improved performance over several baseline models.10 CHAPTER 1. We apply our framework to prediction of disulfide connectivity in proteins using perfect non-bipartite matchings. Learning to match We use combinatorial properties of weighted matchings to develop an exact.

We address the exponential blow-up of the naive problem formulation by deriving two general equivalent convex formulation. THESIS OUTLINE 11 clustering metric for a variety of desired clusterings. We also discuss approximations. Markov networks: We review maximum conditional likelihood estimation and present maximum margin estimation for Markov networks. We define hypothesis classes.5. We describe representation and inference for Markov networks. Structured maximum margin estimation: This chapter outlines the main principles of maximum margin estimation for structured models. We use graphical model decomposition to derive a convex. Chapter 5. We compare probabilistic models. Chapter 4. min-max and certificate. 1. We also briefly describe context free grammars and combinatorial structures as models. compact formulation that seamlessly integrates kernels with graphical models. These formulations. They lead to polynomial size programs for estimation of models where the prediction problem is tractable. We describe the relationship between the dual estimation problems and kernels. risk. conditional likelihood and margin. we define the abstract class of structured prediction problems and models addressed by the thesis. We analyze the theoretical generalization . including email folder organization and image segmentation. loss functions. for solving intractable or very large problems. including dynamic and linear programming inference. and review estimation methods based on maximizing likelihood. including logistic regression and support vector machines. in particular using upper and lower bounds. Supervised learning: We review basic definitions and statistical framework for classification. Chapter 3. allow us to exploit decomposition and combinatorial structure of the prediction task. Structured models: In this chapter.1. generative and discriminative and unnormalized models. We consider generalized linear models.5 Thesis outline Below is a summary of the rest of the chapters in the thesis: Chapter 2.

we provide an approximation that works well in practice.12 CHAPTER 1. Chapter 9. We provide experimental results on a webpage classification task. We present an AMN-based method for object segmentation from 3D range data that scales to very large prediction tasks involving tens of millions of points. Relational Markov networks: We introduce the framework of relational Markov networks (RMNs). where our models significantly outperform other approaches by effectively capturing correlation between adjacent letters and incorporating high-dimensional input representation via kernels. INTRODUCTION properties of max-margin estimation and derive a novel margin-based bound for structured classification. We show . As we show. Associative Markov networks: We define an important subclass of Markov networks. max-margin estimation allows exact training of networks of arbitrary connectivity and topology. associative Markov networks (AMNs). for which maximum likelihood methods are believed to be intractable. Context free grammars: We present max-margin estimation for natural language parsing on the decomposition properties of context free grammars. Chapter 6. called Structured SMO. Chapter 7. M3 N algorithms and experiments: We present an efficient algorithm for solving the estimation problem in graphical models. that captures positive interactions present in many domains. We present experiments with handwriting recognition. and can easily model complex patterns over related entities. Our onlinestyle algorithm uses inference in the model and analytic updates to solve extremely large quadratic problems. For the non-binary case. The graphical structure of an RMN is based on the relational structure of the domain. which compactly defines templates for Markov networks in domains with rich structure modeled by objects. showing that accuracy can be significantly improved by modeling relational dependencies. We show that for associative Markov networks of over binary variables. the use of an undirected. Chapter 8. discriminative graphical model avoids the difficulties of defining a coherent generative model for graph structures in directed models and allows us tremendous flexibility in representing complex patterns. attributes and relations.

Chapter 11. We show experimental evidence of the model’s improved performance over several baseline models. Correlation clustering: In this chapter. efficient algorithm for learning the parameters of the model.6. Experiments on synthetic and real-world data show the ability of the algorithm to learn an appropriate clustering metric for a variety of desired clusterings. The polynomial formulation of maximum margin Markov networks presented in Ch. which makes it possible to efficiently embed the features in very high-dimensional spaces and achieve stateof-the-art accuracy. using a dual decomposition technique [Taskar et al. including email folder organization and image segmentation. Chapter 10. We use combinatorial properties of weighted matchings to develop an exact. Work on associative Markov networks (Ch. Chapter 12..1. 5 was published for a less general case. applicability and limitations.6 Previously published work Some of the work described in this thesis has been published in conference proceedings. 1.. although they underly several papers mentioned below. Perfect matchings: We apply our framework to learning to predict disulfide connectivity in proteins using perfect matchings. Conclusions and future directions: We review the main contributions of the thesis and summarize their significance. We formulate the approximate learning problem as a compact convex program with quadratic objective and linear or positive-semidefinite constraints. The algorithm we propose uses kernels. A paper . 2004a]. 2003a]. We discuss extensions and future research directions not addressed in the thesis. 7) was published with experiments on hypertext and news-wire classification [Taskar et al. The min-max and certificate formulations for structured max-margin estimation have not been published in their general form outlined in Ch. we derive a max-margin formulation for learning affinity scores for correlation clustering from clustered training data. 4. PREVIOUSLY PUBLISHED WORK 13 that this framework allows high-accuracy parsing in cubic time by exploiting novel kinds of lexical information.

which presents a experiments on terrain classification and other tasks. Taskar et al. Natural language parsing in Ch. INTRODUCTION on 3D object segmentation using AMNs. [2003b] defined and applied the Relational Markov networks (Ch. 11. work on correlation clustering in Ch. using maximum (conditional) likelihood estimation. Finally. Dinkar Gupta. [2002] and Taskar et al. 8). Daphne Koller and Andrew Ng). is currently under review (joint work with Drago Anguelov. 9 was published in Taskar et al. [2004b]. Vassil Chatalbashev. done jointly with Pieter Abbeel and Andrew Ng. . 10 (joint work with Vassil Chatalbashev and Daphne Koller) is currently under review. Disulfide connectivity prediction using perfect matchings in Ch.14 CHAPTER 1. Geremy Heitz. has not been published.

Part I Models and methods 15 .

for in-depth discussion of these and many other models). . The second crucial dimension of a learning algorithm is the criterion for selection of h from H. . (independent and identically distributed) samples S = {(x(i) . we seek a function h : X → Y that maps inputs x ∈ X to outputs y ∈ Y. and extensibility to more complex structured prediction tasks will consider in the next chapter. .d. The input to an algorithm is training data. [2000]. including decision trees.1). . where k is the number of classes. generalized log-linear models and kernel methods (see Quinlan [2001]. is called classification. We will concentrate on the last two classes. chief among them is the hypothesis class H of functions h the algorithm outputs. Duda et al. The goal of a learning algorithm is to output a hypothesis h such that h(x) will approximate y on new samples from the distribution (x. y) ∼ D. Hastie et al. nearest-neighbors. a set of m i. including accuracy.i. The input space X is an arbitrary set (often X = IRn ). Bishop [1995]. y (i) )}m drawn from a fixed but unknown distribution D i=1 over X × Y. In handwritten character recognition. [2001]. yk }.Chapter 2 Supervised learning In supervised learning. neural networks. 2. while the output space Y we consider in this chapter discrete. We arrive at such a criterion by quantifying what it means for h(x) to approximate 16 . for example. for several reasons we discuss below. Y = {y1 . efficiency. Numerous classes of functions have been well studied. X is the set of images of letters and Y is the alphabet (see Fig. Learning algorithms can be distinguished among several dimensions. A supervised learning problem with discrete outputs.

For a vast majority of supervised learning algorithms. (2. we can always find a polynomial h that passes through all the sample points (x(i) . the degree of the polynomial) with the error on the training data as measured by empirical risk RS . h(x)) ≡ 1 = h(x)). RS [h] is often called the training If our set of hypotheses. In general. y. For 0/1 loss. However. h(x))]. h(x )) = m (i) (i) (i) (i) i (h(x )). y. I(y where 1 denotes the indicator function. y (i) . that is. Since we do not generally know the distribution D. computed on the training sample S: 1 RS [h] = m m i=1 1 (x . y . that is. For example. if X = IR..1) : X × Y × Y → IR+ measures the penalty for predicting h(x) on the sample (x. m i (h(x i=1 (i) )).17 y. m) assuming that all the x(i) are unique. . simply selecting a hypothesis with lowest risk h∗ = arg min RS [h]. it will have zero empirical risk. where the loss function (2. y). This polynomial is very likely to overfit the training data.2) where we abbreviate (x(i) . h(x(i) )) = error or training loss.g. is large enough. h∈H is generally not a good idea. ˆ ˆ A common loss function for classification is 0/1 loss 0/1 (x. this fundamental balance is . i = (1. RS [h] is simply the proportion of training examples that h misclassifies.. y (i) ).y)∼D [ (x. we will be able to find h that has zero or very small empirical risk. we assume that (x.. The key to selecting a good hypothesis is to trade-off complexity of class H (e. Y = IR and H includes all polynomials of degree m − 1. y. we estimate the risk of h using its empirical risk RS . y ) = 0 if y = y . 1 I(·) I(true) = 1 and 1 I(false) = 0. but high actual risk. H. The risk functional RD [(h)] measures the expected error of the approximation: RD [h] = E(x.

and the regularization D[h]. Hastie et al. The formulation in (2. 2001]. y∈Y (2. the search for the optimal h∗ in (2. Below however. and C ≥ 0 is a tradeoff parameter. it is intractable [Bishop. SUPERVISED LEARNING achieved by minimizing the weighted combination of the two criteria: h∗ = arg min D[h] + CRS [h] .3) may be a daunting task1 . We assume that ties in the arg max are broken using some arbitrary but fixed rule. As we discuss below. y). 1995.3) where D[h] measures the inherent dimension or complexity of h. the search procedure used by the learning algorithm is crucial. For these intractable classes. For many classes. we will concentrate on models where the optimal h∗ can be found efficiently using convex optimization in polynomial time. but simply adopt the standard measures as needed and refer the reader to Vapnik [1995]. In general. x. x. we consider hypothesis classes of the following parametric form: hw (x) = arg max f (w. this class of hypotheses is very rich and includes many standard models. where w ∈ W is a set of parameters.4) where f (w. Depending on the complexity of the class H. [2001] for details.4) of the hypothesis class in terms of an optimization procedure will become crucial to extending supervised learning techniques to cases where the output space Y is more complex. 1 . Hence. For classification. [1996]. See discussion in the next section about approaches to dealing with 0/1 loss. y) is a function f : W × X × Y → IR. and we must resort to approximate. The term D[h] is often called regularization. for example decision trees and multi-layer neural networks. the learning algorithms we consider are completely characterized by the hypothesis class H. greedy optimization methods. minimizing the objective with the usual 0/1 training error is generally a very difficult problem with multiple maxima for most realistic H. Quinlan. Devroye et al. We will not go into derivation of various complexity measures D[h] here. usually with W ⊆ IRn . the loss function .18 CHAPTER 2. h∈H (2.

2. differ primarily in their choice of the upper bound on the training 0/1 loss. The standard solution is minimizing an upper bound on the 0/1 loss. y) = arg max w f (x. col) and char ∈ Y.1: Handwritten character recognition: sample letters from Kassel [1995] data set. logistic regression and support vector machines. (x. . Intuitively. y∈Y (2. h(x)) ≥ (x.2. a zero.1 Classification with generalized linear models For classification. we consider the generalized linear family of hypotheses H.. h(x)).1. y) = 1 row. 2. Given n real-valued basis functions fj : X × Y → IR. y. . We might have a basis function fj (x. y) and hence is simpler than a hw with a large weight wj . or small weight wj implies that the hypothesis hw does not depend on the value of fj (x.col = on ∧ y = char) for each possible (row.5) Consider the character recognition example in Fig. The most common loss for classification is 0/1 loss. Our input x is a vector of pixel values of the image and y is the alphabet {a. z}. y). CLASSIFICATION WITH GENERALIZED LINEAR MODELS 19 a b c d e Figure 2. this very simple model captures enough information to perform reasonably well. col). a hypothesis hw ∈ H is defined by a set of n coefficients wj ∈ IR such that: n hw (x) = arg max y∈Y i=1 wj fj (x. . . there are statistical benefits of minimizing a convex upper bound [Bartlett et al. I(x where xrow. The regularization D[hw ] for the linear family is typically the norm of the parameters ||w||p for p = 1. . Two of the primary classification methods we consider. y. 2003]). Minimizing the 0/1 risk is generally a very difficult problem with multiple maxima for any large class H.col denotes the value of pixel (row. 2.1. (In addition to computational advantages of this approach. Since different letters tend to have different pixels turned on.

where y is the correct label for x.2: 0/1-loss upper bounded by log-loss and hinge-loss.2).6) y∈Y exp{w f (x. SUPERVISED LEARNING 3.7) where C is a user-specified constant the determines the trade-off between regularization and likelihood of the data. y)}. we use 2-norm below: min 1 ||w||2 + C 2 log Zw (x(i) ) − w f (x(i) .5 0 −2 −1. i (2. 2. y ). The optimal weights are selected by maximiz- ing the conditional likelihood of the data (minimizing the log-loss) with some regularization.5 −1 −0. we assign a probabilistic interpretation to the hypothesis hw as defining a conditional distribution: Pw (y | x) = where Zw (x) = 1 exp{w f (x.5 1 0. 2. Common choices for regularization are 1 or 2-norm regularization on the weights. while the vertical axis show the value of the associated loss.5 2 Figure 2. This approach is called the (regularized) maximum likelihood estimation. y (i) ). The log-loss is shown up to an additive constant for illustration purposes.5 1 1. Zw (x) (2. Horizontal axis shows w f (x. y)}.5 2 1.2 Logistic regression In logistic regression.5 log−loss hinge−loss 0/1−loss 3 2. . The log-loss log Zw (x) − w f (x.5 0 0.20 CHAPTER 2. y) − maxy =y w f (x. y) is an upper bound (up to a constant) on the 0/1 loss 0/1 (see Fig.

3. We can interpret logistic regression as trying to match the empirical basis function expectations while maintaining a high entropy conditional distribution Pw (y | x). 1998. The gradient with respect to w is given by: w+C i Ei. 2.w [fi (x(i) .8) Ei. y (i) ) = w − C i Ei.w [∆fi (y)] = 0. ||w|| y=y(i) ||w|| y=y(i) . − i. y) = min w ∆fi (y).w [∆fi (y)]. y)] − fi (x(i) .4 Support vector machines Support vector machines [Vapnik. ∀i. the margin on example i quantifies by how much the true label “wins” over the wrong ones: γi = 1 1 min w f (x(i) .2. LOGISTIC DUAL AND MAXIMUM ENTROPY 21 2. y (i) ) − f (x(i) . where Ei. 1991] that the dual of the maximum likelihood problem (without regularization) is the maximum entropy problem: max s. Crammer & Singer. In the multi-class SVM formulation [Weston & Watkins. It can be shown [Cover & Thomas. y (i) ) − w f (x(i) . 2001]. Ignoring the regularization term. so we have an unconstrained (differentiable) convex optimization problem. the gradient is zero when the basis function expectations are equal to the basis functions evaluated on the labels y (i) . y).w [f (y)] = y f (y)Pw (y | x(i) ) is the expectation under the conditional distribu- tion Pw (y | x(i) ) and ∆fi (y) = f (x(i) .t.y Pw (y | x(i) ) log Pw (y | x(i) ) (2.3 Logistic dual and maximum entropy The objective function is convex in the parameters w. 1995] select the weights based on the “margin” of confidence of hw .

9) ∀i. ∀i. We can also express ξi as maxy 0/1 i (y) − w ∆fi (y). (2. ∀i.9). max i. (y) − ξi . y. (2. y (i) ) − f (x(i) .5 SVM dual and kernels The form of the dual of Eq. The solution to the dual α gives the solution to the primal as a weighted combination of basis functions of examples: w∗ = i.y (2. y αi (y) = C.y ∗ αi (y)∆fi (y).2). 0/1 In the dual. y). (2. 2. (2. Maximizing the smallest such margin (and allowing for negative margins) is equivalent to solving the following quadratic program: min s. the αi (y) variables correspond to the w ∆fi (y) ≥ ∗ (y)−ξi constraints in the primal Eq.9) in a form similar to Eq. Note that the slack variable ξi is constrained to be positive in the above program since w ∆fi (y (i) ) = 0 and 0/1 (y (i) ) = 0.11) s. .y αi (y) 0/1 i (y) 1 − 2 2 αi (y)∆fi (y) i. (2.7): min 1 ||w||2 + C 2 max[ i y 0/1 i (y) − w ∆fi (y)].9) is crucial to efficient solution of SVM and the ability to use a high or even infinite dimensional set of basis functions via kernels.t.22 CHAPTER 2. 1 ||w||2 + C 2 w ∆fi (y) ≥ ξi i 0/1 (2. αi (y) ≥ 0. SUPERVISED LEARNING where ∆fi (y) = f (x(i) . 0/1 i (y) − w ∆fi (y)] is also an upper bound on the 0/1 loss 0/1 2. and the optimization problem Eq. ∀y ∈ Y.10) The hinge-loss maxy [ (see Fig.t.

cold = on ∧ y = char) for each I(x possible (row1 . ∧ xrowd .y j. logistic regression can also be kernelized. Computing this polynomial kernel can be done independently of the dimension d. we can x ¯ solve Eq.. Note that at classification time. the hinge loss formulation usually produces sparse solutions in terms of the number of support vectors.2. An important feature of the dual formulation is that the basis functions f appear only as dot products. y) f (x. . (rowd . Zhu & Hastie.y = i. . (2. while solutions to the corresponding kernelized log-loss problem are generally non-sparse (all examples are support vectors) and require approximations for even relatively small datasets [Wahba et al. Expanding the quadratic term. even though the number of basis functions grows exponentially with d [Vapnik. ¯ ¯ For example. we might have basis functions that are polynomial of degree d in terms of image pixels. y (i) ) f (x. are called support vectors. as long as the dot product f (x. 1995]. y y Hence. 2001]. y). However.¯ y αi (y)αj (¯)∆fi (y) ∆fj (¯).col1 = on ∧ .y αi (y)∆fi (y) f (x. (i. y) = 1 row1 . we also do not need to worry about the dimension of f since: w f (x. y )−f (x(i). fj (x.5.11) independently of the actual dimension of f . that have non-zero αi (y). y ) = ¯ i. y) f (¯ . cold ) and char ∈ Y. . we have: 2 αi (y)∆fi (y) i. . y ) = ¯ i. y )]. . col1 ) . SVM DUAL AND KERNELS 23 ∗ The pairings of examples and incorrect labels. 1993. y ) can be computed efficiently. In fact.y αi (y)[f (x(i).

. e. We refer to joint spaces with constraints and correlations as structured. . the output variables are often highly correlated. consecutive letters in a word. . machine vision.Chapter 3 Structured models Consider once more the problem of character recognition. each Yj is the alphabet. In word recognition. In addition to “hard” constraints. . Structured models we consider in this chapter (and thesis) predict the outputs jointly. yL ).g. respecting the constraints and exploiting the correlations in the output space. but in the context of the surrounding letters that together form a word. In fact. for example an entire sequence of L characters.” Distinguishing between the second letter and fourth letter (‘r’ and ‘c’) in isolation is far from trivial. we might restrict that the letter ‘q’ never follows by ‘z’ in English. . while Y is the dictionary. The range of prediction problems these broad definitions encompass is immense. .1 shows an example handwritten word “brace. 3. × YL we consider is a subset of product of output spaces of single variables. we consider prediction problems in which the output is not a single discrete value y. but a set of values y = (y1 . The output space Y ⊆ Y1 × . Fig. Y ⊂ Y1 ×. a more natural and useful task is recognizing words and entire sentences. In this chapter. We call problems with discrete output spaces structured classification or structured prediction.×YL . This joint output space is often a proper subset of the product of singleton output spaces. For concreteness. and computational 24 . In word recognition. let the number of variables L be fixed. . this task is much less error-prone for humans and should be for computers as well. arising in fields as diverse as natural language analysis. .

This formulation is very general. we use an approximate polynomial time optimization procedure. convex optimization such as linear. Clearly. where g : X × Y → IRk . combinatorial optimization problems like min-cut and matching. we will restrict our attention to models where this optimization problem can be solved in polynomial time.3. y) to the input and output space X × Y with p(x.1) where as before f (x. PROBABILISTIC MODELS: GENERATIVE AND CONDITIONAL 25 b r a c e Figure 3. The class of structured models H we consider is essentially of the same form as in previous chapter. for example. 11). In other cases. For the most part. y) ≤ 0} is defined using a vector of functions g(x. for many f . g pairs. finding the optimal y is intractable. The output space Y = {y : g(x. y). biology.y)≤0 (3. quadratic and semi-definite programming. A generative model assigns a normalized joint density p(x. y) that define the constraints. 3. y∈Y x∈X p(x.1: Handwritten word recognition: sample from Kassel [1995] data set. . y) ≥ 0. except that y has been replaced by y: hw (x) = arg max w f (x. to name a few. y) is a vector of functions f : X × Y → IRn .1. y) = 1. probabilistic models like Markov networks (in certain cases) and context-free grammars. 8) and correlation clustering (Ch. like intractable Markov networks (Ch. which can be subdivided into generative and conditional with respect to the prediction task.1 Probabilistic models: generative and conditional The term model is often reserved for probabilistic models. This includes. y : g(x.

2001]. we may be tuning the approximation away from optimal conditional distribution p(y | x). however. Of course. A generative model must make simplifying assumptions (more precisely. 1970. In addition. which we use to make the predictions. y) will not necessarily do so. Hastie et al.. Because of this. the maximum likelihood parameters w with respect to the joint distribution can be computed in closed form using the empirical basis function expectations ES [f (x. . while a conditional model makes many fewer assumption by focusing on p(y | y). y)] [DeGroot. this efficiency comes at a price. generative models are typically structured to allow very efficient maximum likelihood learning. STRUCTURED MODELS A conditional model assigns a normalized density p(y | x) only over the output space Y with p(y | x) ≥ 0. the conditional model will learn the best approximation to p(y | x) possible using w. by optimizing the model to fit the joint distribution p(x. Any model is an approximation to the true distribution underlying the data. y). independence assumptions) about the entire p(x. y)}. y∈Y p(y | x) = 1 ∀x ∈ X . y) ∝ exp{w f (x. generative models actually need fewer samples to converge to a good estimate of the joint distribution than conditional models need to accurately represent the conditional distribution. Given sufficient data. Probabilistic interpretation of the model offers well-understood semantics and an immense toolbox of methods for inference and learning. generative models may actually outperform conditional models [Ng & Jordan. For exponential families. y). A very common class of generative models is the exponential family: p(x. 2001]. while the generative model p(x. It also provides an intuitive measure of confidence in the predictions of a model in terms of conditional probabilities.26 CHAPTER 3. In a regime with very few training samples (relative to the number of parameters w). Typically.

Even more importantly. which do not represent a conditional distribution. we use the term model very broadly. but concentrate on the margin or decision boundary. We still heavily rely on the representation and inference tools familiar from probabilistic models for the construction of and prediction in unnormalized models.2 Prediction models: normalized and unnormalized Probabilistic semantics are certainly not necessary for a good predictive model if we are simply interested in the optimal prediction (the arg max in Eq.1)). This fact makes standard maximum likelihood estimation infeasible. 1995. The learning methods we advocate in this thesis circumvent this problem by requiring only the maximization problem to be tractable. support vector machines.3 Markov networks Markov networks provide a framework for a rich family of models for both discrete and continuous prediction [Pearl.2. 3. Cowell et al. while the optimal y can be found in polynomial time. the difference between the optimal y and the rest. 2000].. As we discussed in the previous chapter. (3. typically perform as well or better than logistic regression [Vapnik. 1999]. Cristianini & Shawe-Taylor. but largely dispense with the probabilistic interpretation when needed. In this chapter. Essentially. In general. The models treat the inputs and outputs as random variables X with domain X and Y with domain Y and compactly define a conditional density p(Y | X) or distribution P (Y | X) (we concentrate here on the . we review basic concepts in probabilistic graphical models called Markov networks or Markov random fields.3. in many cases we discuss below. We also briefly touch upon examples of context-free grammars and combinatorial problems that will be explained in greater detail in Part III to illustrate the range of prediction problems we address. normalizing the model (summing over the entire Y) is intractable. 1988. PREDICTION MODELS: NORMALIZED AND UNNORMALIZED 27 3. we can often achieve higher accuracy models when we do not learn a normalized distribution over the outputs. to include any scheme that assigns scores to the output space Y and has a procedure for finding the optimal scoring y.

where the nodes are associated with variables V = {Y1 . . E). . x). 3. the correlation between a letter Yb before j and a letter Ya after j is negligible. {Y1 . . No assumption is . x). Y2 }. We do not explicitly represent the inputs X in the figure. A clique is a set of nodes c ⊆ V that form a fully connected subgraph (every two nodes are connected by an edge).3. The quantitative aspect of the model is defined by the potentials that are associated with nodes and cliques of the graph. {Y5 }. the model encodes that Yj is conditionally independent of the rest of the variables given Yj−1 . STRUCTURED MODELS conditional Markov networks or CRFs [Lafferty et al. Y5 }}. YL }. . In the chain network in Fig. For example.2. and the optimal subsequence starting with yj from j to L.2. . More precisely. the cliques are simply the nodes and the edges: C(G) = {{Y1 }.1 Representation The structure of a Markov network is defined by an undirected graph G = (V. 2001]). . Yj+1 . b < j < a. . but many of the same representation and inference concepts translate to continuous domains. For the purposes of finding the most likely y. x) = P (Yb | Yj . and we consider each node a singleton clique. 3. {Y4 . an assignment of variables in the clique as yc and the space of all assignments to the clique as Yc . . it suffices to separately find the optimal subsequence from 1 to j ending with yj . x) = P (Ya | Yj . 3. Yb . We focus on discrete output spaces Y below. Note that each subclique of a clique is also a clique. . . adjacent letters in a word are highly correlated.28 CHAPTER 3. .. . the nodes are associated with output variables Yi and the edges correspond to direct dependencies or correlations. we use a model where P (Yb | Yj . but the first-order model is making the assertion (which is certainly an approximation) that once the value of a letter Yj is known. The graphical structure of the models encodes the qualitative aspects of the distribution: direct dependencies as well as conditional independencies. In Fig. consider a first-order Markov chain a model for the word recognition task. Intuitively. this conditional independence property means that the optimization problem is decomposable: given that Yj = yj . P (Ya | Yj . Ya . We denote the set of variables in a clique c as Yc . Before a formal definition. The advantage of a graphical framework is that it can exploit sparseness in the correlations between outputs Y .

made about X . MARKOV NETWORKS 29 Figure 3. . The nodes are associated with variables V = {Y1 . yc ) with φc : X × Yc → IR+ .i+1 (Yi . Potentials do not have a local probabilistic interpretation. The Markov network (G. we have node and edge potentials. For simplicity.3.‘e’). E) and a set of potentials Φ = {φc }. . yc ). assume that the edge potentials would not depend . but can be thought of as defining an unnormalized score for each assignment in the clique. Each clique c ∈ C(G) is associated with a potential φc (x.3. In our example Fig. φi.‘r’. 3. Intuitively. which specifies a non-negative value for each assignment yc to variables in Yc and any input x. .2. Conditioned on the image input. though perhaps there would be some ambiguity with the second and fourth letter.‘a’. Definition 3.3.‘c’. Yi+1 ) are edge potentials (dependence on x is not shown). yc ). Φ) defines a conditional distribution: P (y | x) = 1 Z(x) φc (x. .1 A Markov network is defined by an undirected graph G = (V. the node potentials quantify the correlation between the input x and the value of the node. appropriate node potentials in our network should give high scores to the correct letters (‘b’. while the edge potentials quantify the correlation between the pair of adjacent output variables as well as the input x.2: First-order Markov chain: φi (Yi ) are node potentials. YL }. c∈C(G) where C(G) is the set of all the cliques of the graph and Z(x) is the partition function given by Z(x) = y∈Y c∈C(G) φc (x.

we condition a clique only on a portion of the input x. yc ) − log Zw (x). For the edge potentials.col = on ∧ yj = char) for each I(x possible (row. all single node potentials would share the same weights and basis functions (albeit the relevant portion of the input xc is different) and similarly for the pairwise cliques. col) in xj .j+1. so when c is a node. yj . we stack all basis functions into one vector f . these scores should favor the correct output “brace”. Hence the log of the conditional probability is given by: log P (y | x) = c∈C(G) wc fc (x. when c is an edge. STRUCTURED MODELS on the images.row. yc ) are defined to evaluate to zero. yc ) = exp k=1 wc. In fact.k (x. yc ) where nc is the number of basis functions for the clique c.30 CHAPTER 3.1 With slight abuse of notation. In general. yc ) could be represented (in log-space) as a sum of basis functions over x. yc ) = exp wc fc (x. since the potentials φc (xc . For the sequence model. but simply should give high scores to pairs of letters that tend to appear often consecutively. a Markov network is a generalized log-linear model.k (x. the window of the image that corresponds to letter j and each char ∈ Yj (we assume the input has been segmented into images xj that correspond to letters). Similarly. no matter in what position they appear in the sequence. which we denote as xc . In case of node potentials for word recognition. we could use the same basis functions as for individual character recognition: fj. we are likely to “tie” or “share” the parameters of the model wc across cliques. the edge functions in f (xc . yc : nc φc (xc . which can be accomplished by adding more basis functions that condition on the position of the clique. the node Sometimes we might actually want some dependence on the position in the sequence.k fc. 1 . Multiplied together. yj ) = 1 j. Usually. we can define basis functions for each combination of letters (assume for simplicity no dependence on x) : fj. f has node functions and edge functions. yj+1 ) = 1 j = char1 ∧yj+1 = I(y char2 ) for each char1 ∈ Yj and char2 ∈ Yj+1 . In this problem (as well as many others).k (x.

yc ) are also defined to evaluate to zero. is just one of such questions. so the most likely assignment according to the model is given by: arg max log Pw (y | x) = arg max w f (x. known as maximum a-posteriori (MAP) or most likely explanation (MPE). 1 < k ≤ L. say according to some lexicographic order of the symbols. The Viterbi dynamic programming algorithm solves this problem for chain networks in O(L) time. The task of finding the most likely assignment.1).k−1 φj (x. we can back-trace the highest scoring sequence itself.3. max φ∗ (yk−1 )φj (x. yc ). Now we can write: f (x. ∀yk ∈ Yk . yj )φj (x.2 Inference There are several important questions that can be answered by probabilistic models. y1 ). 3. yk )φj (x..3. Using the arg max’s of the max’s in L the computation of φ∗ . yj−1 . MARKOV NETWORKS 31 functions in f (xc . yj ). y∈Y y∈Y in the same form as Eq. y). We stack the weights in the corresponding manner. k−1 The highest scoring sequence has score maxyL φ∗ (yL ). j The algorithm computes the highest scores recursively: φ∗ (y1 ) = φ1 (x. (3. but most relevant to our discussion.3. 1 φ∗ (yk ) = k yk−1 ∈Yk−1 ∀y1 ∈ Y1 . y) = c∈C(G) f (xc . Let the highest score of any subsequence from 1 to k > 1 ending with value yk be defined as φ∗ (yk ) = max k y1. yk ). We assume that score ties are broken in a predetermined way. yk−1 . .

3. there are several important subclasses of networks that allow polynomial time inference. Y3 ) or (Y2 . . In general. MAP inference is NP-hard [Cowell et al. there are many possible sets of edges that can be added to triangulate a graph. The tree-width of an untriangulated graph G is the minimum tree-width of all triangulations of G.2 (Triangulated graph) An undirected graph G is triangulated if every one of its cycles of length ≥ 4 possesses a chord.32 CHAPTER 3.3..3: Diamond Markov network (added triangulation edge is dashed). The critical property of a triangulation for the inference procedure is the size of the largest clique. The most important of these is the class of networks with low tree-width. STRUCTURED MODELS Figure 3. which are connected by edges (vi . The inference procedure creates a tree of cliques using the graph augmented by triangulation. . 1999]. v1 . . . vl ). . Singly-connected graphs. are triangulated since they contain no cycles. Y4 ). To triangulate it. The simplest untriangulated network is the diamond in Fig. However. Recall that a cycle of length l in an undirected graph G is a sequence of nodes (v0 . We need the concept of triangulation (or chordality) to formally define tree-width. we can add the edge (Y1 . vj ) ∈ G between non-consecutive nodes. In general Markov networks. distinct except that v0 = vl .3. Definition 3. Definition 3.3 (Tree-width of a graph) The tree-width of a triangulated graph G is the size of its largest clique minus 1.3. vi+1 ) ∈ G. A chord of this cycle is an edge (vi . like chains and trees.

Each of the original clique potentials must associated with exactly one node in the junction tree. Definition 3. ∀yc ∈ Yc .3. Y2 . E) for a triangulated graph G is a tree in which the nodes are a subset of the cliques of the graph. Fig. ∀yl ∈ Yl . Y3 .3. but good heuristic algorithms exist [Cowell et al. Y3 ) = φ2 (Y2 )φ3 (Y3 )φ12 (Y1 . yc ) c c ∈Chr (c) ∀l ∈ Lv(T ). Y4 )φ34 (Y3 . 3. The inference procedure is based on a data structure called junction tree that can be constructed for a triangulated graph. We denote the leaves of the tree as Lv(T ) and the children of node c (relative to the root r) as Chr (c): φ∗ (yl ) = φl (x. Y4 } nodes are the product of the associated clique potentials: φ134 (Y1 . The Viterbi algorithm for junction trees picks an arbitrary root r for the tree T and proceeds recursively from the leaves to compute the highest scoring subtree at a node by combining the subtrees with highest score from its children. 1999]. For example. Y2 )φ23 (Y2 .. V ⊆ C(G) and the edges E satisfy the running intersection property: for any two cliques c and c . [1999]. Y3 . l φ∗ (yc ) = φc (x. c ∀c ∈ V(T ) \ Lv(T ). Algorithms for constructing junction trees from triangulated graphs are described in detail in Cowell et al. .4 (Junction tree) A junction tree T = (V. Y4 } and {Y1 . 3. yl ). the potentials for the {Y1 .3. Finding the minimum tree-width triangulation of a general graph is NP-hard. Y4 ). the variables in the intersection c ∩ c are contained in the clique of every node of the tree on the (unique) path between c and c .3 is 2. Y3 ). Y3 . yc ∼yc max φ∗ (yc ). The junction tree is an alternative representation of the same distribution that allows simple dynamic programming inference similar to the Viterbi algorithm for chains.4 shows a junction tree for the diamond network. MARKOV NETWORKS 33 The tree-width of a chain or a tree is 1 and the tree-width of Fig. φ123 (Y1 . Y4 ) = φ1 (Y1 )φ4 (Y4 )φ14 (Y1 .

4: Diamond network junction tree. we present an alternative inference method based on linear programming. yc ) = c. this formulation is crucial in viewing Markov networks in a unified framework of the structured models we consider and to our development of common estimation methods in later chapters. . The highest score is given by maxyr φ∗ (yr ).. such that: log c φc (x.3 Linear programming MAP inference In this section. Note that this algorithm is exponential in the tree-width. the size of the largest clique. Using the arg max’s of the max’s in the computation of φ∗ .34 CHAPTER 3. Similar type of computations using the junction tree can be used to compute the partition function Zw (x) (by simply replacing max by ) as well as marginal probabilities P (yc |x) for the cliques of the graph [Cowell et al. 1999]. Although solving the MAP inference using a general LP solver is less efficient than the dynamic programming algorithms above. 3. where yc ∼ yc denotes whether the partial assignment yc is consistent with the partial assignment yc on the variables in the intersection of c and c . Each of the original potentials is associated with a node in the tree.yc µc (yc ) log φc (x. we can backr trace the highest scoring assignment itself. STRUCTURED MODELS Figure 3. yc ). one for each clique c and each value of the clique yc . that denotes whether the assignment has that value. We represent an assignment as a set binary variables µc (yc ).3. Let us begin with a linear integer program to compute the optimal assignment y.

∀yc . We call these variables marginals. There are several elementary constraints that such marginals satisfy. 1}. ∀c ∈ C. for any assignment y . yc ) µc (yc ) = 1.5: Example of marginal agreement: row sums of µ12 (y1 . s. µc (yc ).3. column sums agree with µ2 (y2 ). . y2 ) agree with µ1 (y1 ). For any clique c ∈ C and a subclique s ⊂ c.t. the assignment of the subclique. s ⊂ c. c ∈ C. First. µs (ys ).yc µc (yc ) log φc (x. yc (3. For example. the marginals for cliques that share variables are consistent. ∀s. they must sum to one for each clique. these constraints define a linear integer program: max c. ∀ys .5. Clearly. we can define µc (yc ) variables that satisfy the above constraints by setting µc (yc ) = 1 c = yc ).3. as they correspond to the marginals of a distribution that has all of its mass centered on the MAP instantiation (assuming it is unique). we will have node and edge marginals that sum to 1 and agree with each other as in Fig. must be consistent with the assignment of the clique. MARKOV NETWORKS 35 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 Figure 3. We can also show that converse is true: any I(y valid setting of µc (yc ) corresponds to a valid assignment y.2) µc (yc ) ∈ {0. In fact. in case the network is a chain or a tree. ∀c ∈ C. 3. Second. µs (ys ) = yc ∼ys µc (yc ). Together.

1) = 0. 0) = µ12 (0. µ34 (1. . Note that the edge marginals for (1. 1) = 0. In case the network is not triangulated.5. 0) = µ23 (1. the above marginals will be disallowed. µ12 (1. µ23 (0.2) can be relaxed and the resulting LP is guaranteed to have integer solutions. 0) = µ34 (1. 4) disallow any assignment other than 0000 or 1111.3 with binary variables. the set of marginals is not guaranteed to represent a valid distribution. 3. we discuss another subclass of networks where MAP inference using LPs is tractable for any network topology. 1) = 0. with the following edge marginals that are consistent with the constraints: µ12 (0. Y3 } and {Y1 . 0) = µ14 (0. We can resort to the above LP without triangulation as an approximate inference procedure (augmented with some procedure for rounding possibly fractional solutions). 0) = µ34 (0.5. (2. the value of the LP is the same as the value of the integer program. Y3 . 1) = 0. µ23 (1. (3.5 For a triangulated network with unique MAP assignment.5. 0) = µ23 (0. Hence this set of marginals disallows all assignments. If we triangulate the graph and add the cliques {Y1 .5. 7. 0) = µ12 (1. Consider. In Ch. Y2 . but with a restricted type of potentials. µ14 (0. The optimal distribution with the respect to the objective puts all its mass on the MAP assignment. the constraints force the marginals µc (yc ) to correspond to some valid joint distribution over the assignments. but the edge marginal for (1.3. 4) disallows any assignment that has Y1 = Y4 . [2002]. the diamond network in Fig. Y4 } with their corresponding constraints. 3). the integrality constraint in the integer program in Eq. 0) = µ34 (1. 1) = 0.36 CHAPTER 3. STRUCTURED MODELS Lemma 3. 2). 1) = 0. The corresponding node marginals must all be set to 0. (3. A proof of this lemma appears in Wainwright et al. exact inference is intractable. and any linear combination of the MAP assignments maximizes the LP.5. If the MAP assignment is not unique. for example. Intuitively. 1) = 0. µ34 (0. 1) = 0. In graphs where triangulation produces very large cliques. µ14 (1.

3.1 (CFG) A CFG G consists of: ◦ A set of non-terminal symbols. CONTEXT FREE GRAMMARS 37 Figure 3. .6: Example parse tree from Penn Treebank [Marcus et al. determiners (DT) and prepositions (IN). For example. we restrict our grammars to be in Chomsky normal form2 (CNF).. verbs (VBD)..4 Context free grammars Context-free grammars are one of the primary formalisms for capturing the recursive structure of syntactic constructions [Manning & Sch¨ tze. verbal phrase (VP) or prepositional phrase (PP) and part-of-speech tags like nouns (NN). 1993].4. a primary linguistic resource for expert-annotated English text. N ◦ A designated set of start symbols. B and C are non-terminal symbols. 1999]. This tree is from the Penn Treebank [Marcus et al. Definition 3. 3. 1993]. For clarity of presentation. 3. Fig. where all rules in the grammar are of the form: A → B C and A → D. The terminal symbols (leaves) are the words of the sentence. NS ⊆ N 2 Any CFG can be represented by another CFG in CNF that generates the same set of sentences. and D is a terminal symbol.6 shows u a parse tree for the sentence The screen was a sea of red.4. where A. The non-terminal symbols (labels of internal nodes) correspond to syntactic categories such as noun phrase (NP).

VP. 3. NP. VP → VP PP. NN → cat. NP → NP PP. NN. divided into Binary productions. B. since it has several recursive productions. VP → VBD NP. D ∈ T }. saw. the. C ∈ N } and Unary productions. PU }. An unlikely but possible alternative is that the cat . dog. then rewriting the VP as VP → VBD NP with VBD → saw. For example. DT → the. PU = {A → D : A ∈ N . T ◦ A set of productions. tree. NN → dog. IN → from} A grammar generates a sentence by starting with a symbol in NS and applying the productions in P to rewrite nonterminal symbols. NN → tree. rewriting the NP as NP → DT NN with DT → The and NN → cat. The likely analysis of the sentence is that the cat. VBD. VBD → saw. again using NP → DT NN. from} ◦ PB = {S → NP VP. sitting in the tree. PP → IN NP}. It can also generate the same sentence several ways. there are exponentially many parse trees that produce a sentence of length l. Consider the sentence The cat saw the dog from the tree. P = {PB .6 or (more compactly) using bracketed expressions like the one below: [[TheDT catNN ]NP [sawVBD [theDT dogNN ]NP ]VP ]S . but now with DT → the and NN → dog.38 CHAPTER 3. PB = {A → B C : A. cat. IN} ◦ NS = {S} ◦ T = {The. STRUCTURED MODELS ◦ A set of terminal symbols. DT. In general. We can represent such derivations using trees like in Fig. Consider a very simple grammar: ◦ N = {S. PP. ◦ PU = {DT → The. NP → DT NN. we can generate The cat saw the dog by starting with S → NP VP. The simple grammar above can generate sentences of arbitrary length. saw the dog.

Hence the logprobability of a tree is a sum of the log-probabilities of its productions.1) as hw (x) = arg max w f (x. In Ch. y). CONTEXT FREE GRAMMARS 39 actually saw the dog who lived near the tree or was tied to it in the past. By letting our basis functions f (x. We also need to assign a probability to the different starting symbols P (A) ∈ NS such that A∈NS P (A) = 1.4. The distribution P (x. we can cast PCFG as a structured linear model (in log space). Standard PCFG parsers use a Viterbi-style algorithm to compute arg maxy P (x. B. called prepositional attachment. Our grammar allows both interpretations.y)≤0 and describe the associated algorithm to compute the highest scoring parse tree y given a sentence x. y : g(x. . ∀A ∈ N . This kind of ambiguity. 9. sawVBD [[theDT dogNN ]NP [fromIN [theDT treeNN ]NP ]PP ]NP . y) as the most likely parse tree for a sentence x. y) is defined by assigning a probability to each production and making sure that the sum of probabilities of all productions starting with a each symbol is 1: P (A → B C) = 1.C:A→B C∈PB D:A→D∈PU P (A → D ) = 1. we will show how to represent a parse tree as an assignment of variables Y with appropriate constraints to express PCFGS (and more generally weighted CFGs) in the form of Eq. with the difference being in the analysis of the top-level VP: [sawVBD [theDT dogNN ]NP ]VP [[fromIN [theDT treeNN ]NP ]PP . y) consist of the counts of the productions and w be their log-probabilities. (3. The probability of a tree is simply the product of probabilities of the productions used in the tree (times the probability of the starting symbol). is very common in many realistic grammars. A standard approach to resolving ambiguity is to use a PCFG to define a joint probability distribution over the space of parse trees Y and sentences X .3.

spanning tree. As a particularly simple and relevant example. Treated abstractly.40 CHAPTER 3. graph-cut. Euclidian distance or temporal latency). constructing the weights is an important and labor-intensive design task.t. 1976. where the weights represent the “expertise” of each reviewer for each paper.. 2001]. such as matchings or graph-cuts or trees).k µjk qjk µjk = R. 1982. bin-packing. scheduling. and that each reviewer be assigned at most P papers. j k (3. routing. The bipartite matching problem can be solved using a combinatorial algorithm or the following linear program: max j. consider modeling the task of assigning reviewers to papers as a maximum weight bipartite matching problem. For each paper and reviewer. and many others [Lawler. edge-cover. Cormen et al. s. a combinatorial space of structures. In some domains. More specifically. 1999].3) µjk ≤ P. A simple scheme could measure the “expertise” as the percent of word . and 0 otherwise. This LP is guaranteed to produce integer solutions (as long as P and R are integers) for any weights q(y) [Nemhauser & Wolsey. network design and many more. STRUCTURED MODELS 3. Our objective is to find an assignment for reviewers to papers that maximizes the total weight. the weights of the objective function in the optimization problem are simple and natural to define (for example. job assignment. but in many others. These problems arise in applications such as resource allocation. The quality of the solution found depends critically on the choice of weights that define the objective. Papadimitriou & Steiglitz.5 Combinatorial problems Many important computational tasks are formulated as combinatorial optimization problems such as the maximum weight bipartite and perfect matching. suppose we would like to have R reviewers per paper. We represent a matching with a set of binary variables yjk that take the value 1 if reviewer j is assigned to paper k. together with a scoring scheme that assigns weights to candidate outputs is a kind of a model. 0 ≤ µjk ≤ 1. we have an a weight qjk indicating the qualification level of reviewer j for evaluating paper k.

However. 1979.5. We can define fd (x. In the next chapter we will show how to learn the parameters w in much the same way we learn the parameters w of a Markov network. 1979]. we will represent the objective q(y) as a weighted combination of a set of basis functions w f (x. which is essentially weighted counting. Similarly. since it requires computing the normalization. However. y) = score qjk is simply qjk = jk yjk 1 I(wordd ∈ xjk ). Note that counting is easier than normalization. we would want to weight certain words much more (words that are relevant to the subject and infrequent). y) is easier than the normalization problem y∈Y exp{w f (x. As usual. y)}. Let xjk denote the intersection of the set of words occurring in webpage(j)∩abstract(k). The space of bipartite matchings illustrates an important property of many structured spaces: the maximization problem arg maxy∈Y w f (x. even simply counting the number of matchings is #P-complete [Valiant. exact maximum likelihood estimation is intractable. Then the d words. The maximum weight bipartite matching can be found in polynomial (cubic) time in the number of nodes in the graph using a combinatorial algorithm. y). Garey & Johnson. the number of times word d was in both wd 1 I(wordd ∈ xjk ). the web page of a reviewer j and the abstract of the paper k. Constructing and tuning the weights for a problem is a difficult and time-consuming process. This fact makes a probabilistic interpretation of the model as a distribution over matchings intractable to compute. .3. just as it is for Markov networks for handwriting recognition. COMBINATORIAL PROBLEMS 41 overlap in the reviewer’s home page and the paper’s abstract. a weighted combination of overlapping the web page of a reviewer and the abstract of the paper that were matched in y.

Chapter 4 Structured maximum margin estimation In the previous chapter.y)≤0 w f (x. First and foremost. A compact problem formulation is polynomial in the description length of the objective and the constraints. we develop methods for finding parameters w i=1 such that: arg max w f (x(i) .1) where we assume that the optimization problem maxy : g(x. y) ≤ 0}. Second. Given a sample S = {(x(i) . ∀i. y). y) can be solved or approximated by a compact convex optimization problem for some convex subset of parameters w ∈ W. we described several important types of structured models of the form: hw (x) = arg max w f (x. y) ≈ y(i) . y : g(x. y∈Y (i) where Y (i) = {y : g(x(i) . Both of these approaches define a convex optimization problem for finding such parameters w.y)≤0 (4. we describe at an abstract level two general approaches to structured estimation that we apply in the rest of the thesis. we can use standard optimization software to solve the problem. we can find globally optimal parameters (with fixed precision) in polynomial time. In this chapter. y(i) )}m . There are several reasons to derive compact convex formulations. Although specialpurpose algorithms that exploit the structure of a particular problem are often much faster 42 .

where maximum likelihood estimation is intractable. we will adopt the hinge upper bound i (h(x(i) )) on the loss function for structured classification inspired by max-margin criterion: i (h(x (i) )) = max [w fi (y) + i (y)] − w fi (y(i) ) ≥ i (h(x(i) )). We provide exact maximum margin solutions for several of these problems (Ch. In prediction problems where the maximization in Eq. 4. In structured problems. a natural loss function is a kind of Hamming distance between y(i) and h(x(i) ): the number of variables predicted incorrectly.1 Max-margin estimation As in the univariate prediction. y∈Y (i) . 10). 6). Our framework applies not only to the standard models typically estimated by probabilistic methods. we consider convex programs that provide only an upper or lower bound on the true solution. we can analyze the generalization performance of the framework without worrying about the actual algorithms used to carry out the optimization and the associated woes of intractable optimization problems: local minima. such as Markov networks and context-free grammars. where we are jointly predicting multiple variables. 4. the availability of off-the-shelf software is very important for quick development and testing of such models. the loss is often not just the simple 0-1 loss or squared error. MAX-MARGIN ESTIMATION 43 (see Ch. we measure the error of approximation using a loss function . We will explore these and more general loss functions in the following chapters. (4.1 Min-max formulation Throughout. etc.1.1. greedy and heuristic methods.1) is intractable. For structured classification. but also to a wide range of “unconventional” predictive models. Third.4. 7 and Ch. We discuss how to use these approximate solutions for approximate learning of parameters. Such models include graph cuts and weighted matchings.

1 ||w||2 + C 2 (i) ξi i y∈Y (i) (4. and fi (y) = f (x(i) . The above formulation is a convex quadratic program in w.t. most often IRn ). Assuming ξi are all zero (say because C is very large). (4. If the loss function is not uniform over all the mistakes y = y(i) . ∀i. 1 ||w||2 + C 2 (i) ξi i (4.2) w fi (y ) + ξi ≥ max [w fi (y) + i (y)]. which is generally exponential in Li . (4. y). but its exponential . scaled by the loss i (y). In Ch. ∀y ∈ Y (i) . STRUCTURED MAXIMUM MARGIN ESTIMATION where as before. h(x(i) )). With this upper bound. This form reveals the “maximum margin” nature of the formulation.10): min s. so minimizing ||w|| maximizes the smallest such margin. Another way to express this problem is using i |Y (i) | linear constraints. i (h(x(i) )) = (x(i) . 5 we analyze the effect of non-uniform loss function (Hamming distance type loss) on generalization. the number of variables in yi .t. the min-max formulation for structured classification problem is analogous to multi-class SVM formulation in Eq.9) and Eq. We can interpret 1 w ||w|| [fi (y(i) ) − fi (y)] as the margin of y(i) over another y ∈ Y (i) . min s.2) is that the constraints have a very unwieldy form. then the constraints make costly mistakes (those with high i (y)) less likely. and show a strong connection between the loss-scaled margin and expected risk of the learned model. but assume this throughout this chapter. For brevity. we did not explicitly include the constraint that the parameters are in some legal convex set (w ∈ W. ∀i. y(i) . i (h(x(i) )) = (x(i) . (2.3) is a standard QP with linear constraints. y(i) . the constraints enforce w fi (y(i) ) − w fi (y) ≥ i (y). The formulation in Eq. since maxy∈Y (i) [w fi (y) + i (y)] is convex in w (maximum of affine functions is a convex function). The problem with Eq. (2.3) w fi (y ) + ξi ≥ w fi (y) + i (y). h(x(i) )). The slack variables ξi allow for violations of the constraints at a cost Cξi .44 CHAPTER 4.

the form of the loss term i (y) is crucial for the loss-augmented inference to remain tractable. (4. (4.4) y∈Y (i) Even if maxy∈Y (i) w fi (y) can be solved in polynomial time using convex optimization. (4. (4. The range of tractable losses will depend strongly on the problem itself (f and Y).4) is.4.2) and transform it to a a more manageable problem.4) is. Assume that we find such a formulation in terms of a set of variables µi . consider the example of the reviewer-assignment problem we discussed in the previous chapter: we would like a bipartite matching with R reviewers per paper and at most P papers per reviewer. Likewise.1. with a concave (in µi ) objective fi (w. since Eq. yjk ) = yjk 1 I(wordd ∈ xjk ). Even within the range of tractable losses. We now return to Eq. We can define a basis function fd (xjk . Let xjk denote the intersection of the set of words occurring in the web page of a reviewer j and the abstract of the paper k. The key to solving Eq. we develop a max-margin formulation that uses Lagrangian duality (see [Boyd & Vandenberghe. MAX-MARGIN ESTIMATION 45 size is in general prohibitive. To make the symbols concrete.5) y∈Y (i) µi :gi (µi )≤0 We call such formulation compact if the number of variables µi and constraints gi (µi ) is polynomial in Li . 2004] for an excellent review) to define a joint. (4. A large part of the development of structured estimation methods in the following chapters is identifying appropriate loss functions for the application and designing convex formulations for the loss-augmented inference. µi ). In the next section. we can assume that it is feasible and bounded if Eq. (4. the number of variables in y(i) . µi ) and subject to convex constraints gi (µi ): max [w fi (y) + i (y)] = max fi (w. compact convex problem for estimating the parameters w. some are more efficiently computable than others. Note that maxµi :gi (µi )≤0 fi (w.2) efficiently is the loss-augmented inference max [w fi (y) + i (y)]. which indicates whether the word d is (i) (i) . Let yjk indicate whether reviewer j is matched to the paper k. Each training sample i consists of a matching of Np papers and Nr reviewers from some previous year. µi ) must be convex in w.

yjk ) + 0/1 i. k µi. j k (i) (i) s.jk ≤ P. f (x.jk ≤ 1. 0 ≤ µi.5). y) = jk [w f (xjk .jk [w fjk −yjk ] µi.jk ≤ 1. y) as an LP and we use a loss function this is linear in the number of mistakes.6) for appropriately defined di . ci . y). 0 ≤ µi. we have a linear program of this form for the loss-augmented inference: di + max (Fi w + ci ) µi s. (i) The last equality follows from the fact that any valid matching for example i has R reviewers for Np papers.jk (yjk )] (i) = RNp + jk yjk [w fjk − yjk ]. µi ) = RNp + and gi (µi ) ≤ 0 ⇔ j i. yjk ).jk (yjk ) = jk (i) 1 jk = yjk ) = RNp − I(y jk (i) yjk yjk .t. If this LP is compact (the number of variables and constraints is polynomial in the number of . fi and gi are affine in µi : fi (w. Combining the two pieces. y) and g(x. Fi . For example.46 CHAPTER 4. µi. (4. we have w f (x(i) . which depend only on x(i) .jk [w fjk − yjk ] µi.3) (without the constant term RNp ): max jk (i) µi.jk ≤ P. the Hamming loss simply counts the number of different edges in the matchings y and y(i) : H i (y) (i) (i) = jk 0/1 i.jk = R. y(i) . In general. bi . Ai . Note that the dependence on w is linear and only in the objective of the LP. (i) In terms of Eq. We abbreviate the vector of all the basis functions for each edge jk as yjk fjk = f (xjk . when we can express maxy∈Y (i) w f (x(i) . µi ≥ 0.jk = R. (i) The loss-augmented inference problem can be then written as an LP in µi similar to Eq. hence RNp − (i) (i) jk yjk yjk represents exactly the number of edges that (i) are different between y and y(i) . (4. We assume that the loss function decomposes over the variables yij .j µi. (3. Ai µi ≤ bi .t. STRUCTURED MAXIMUM MARGIN ESTIMATION in both the web page of a reviewer and the abstract of the paper that are matched in y.

λi ) are convex in both w and λi . one for each constraint function in gi (µi ). (4.t. then we can use it to solve the max-margin estimation problem efficiently by using convex duality. λi ) ≤ 0. (4.7) where λi ≥ 0 is a vector of Lagrange multipliers.6) is di + min bi λi s. The Lagrangian associated with Eq. λi ) and qi (w. 1 ||w||2 + C 2 ξi i qi (w. Since we assume that fi (w. we can combine the minimization over λ with minimization over {w. Ai λi ≥ Fi w + ci .4.9) where hi (w. the dual is polynomial size as well. (We folded λi ≥ 0 into qi (w. µi ) = min max Li. λi ) for brevity. (4. λi ) = fi (w.t. (4. The .2). w fi (y(i) ) + ξi ≥ Moreover. we have strong duality: max fi (w.8) into Eq. the dual of the LP in Eq. µi ) − λi gi (µi ). For example.w (µi . λi ). (4. hi (w. ξ}. λi ≥0 µi µi :gi (µi )≤0 For many forms of f and g.w (µi .w (µi .1. −λi ≤ 0}. (4. µi ) is concave in µi and bounded on the non-empty set µi : gi (µi ) ≤ 0.8) where hi (w. λi ) = di + bi λi and qi (w. λi ) qi (w.t. MAX-MARGIN ESTIMATION 47 label variables). λi ) explicitly as: min s. λi ) ≤ 0 is {Fi w + ci − Ai λi ≤ 0.10) min hi (w.λi )≤0 (4. ∀i.) Since the original problem had polynomial size. we can write the Lagrangian dual minλi ≥0 maxµi Li. Plugging Eq. (4. we get min s. λi ).4) is given by Li. λi ≥ 0.

2 Certificate formulation In the previous section. w fi (y(i) ) + ξi ≥ di + bi λi . For our LP-based example. 1997. however. ∀i. (4.4). λi ≥ 0. Schrijver. ξ} that is optimal. For example.48 CHAPTER 4. leading to a suboptimal solution w. though. In some cases. qi (w. can be solved in polynomial time using combinatorial algorithms. Intuitively.t. The exact form of this program depends strongly on f and g.11) w fi (y(i) ) + ξi ≥ hi (w.1. 1 ||w||2 + C 2 ξi i (4. STRUCTURED MAXIMUM MARGIN ESTIMATION reason for this is that if the right hand side is not at the minimum. we have a QP with linear constraints: min s. min s. 1 ||w||2 + C 2 ξi i (4. but no polynomial formulation is known [Bertsimas & Tsitsiklis. . Hence we have a joint and compact convex optimization program for estimating w. Both of these problems. There are several important combinatorial problems which allow polynomial time solution yet do not have a compact convex optimization formulation. just verifying that a given assignment is optimal is sometimes easier than actually finding it. ∀i. we can find a compact certificate of optimality that guarantees that y(i) = arg maxy [w fi (y) + i (y)] without expressing loss-augmented inference as a compact convex program. λi ). ∀i.12) ∀i. 2003]. Ai λi ≥ Fi w + ci . Optimizing jointly over λ as well will produce a solution to {w.t. ∀i. λi ) ≤ 0. we assumed a compact convex formulation of the loss-augmented max in Eq. maximum weight perfect (non-bipartite) matching and spanning tree problems can be expressed as linear programs with exponentially many constraints. the constraint is tighter than necessary. 4.

2) may be infeasible. j k s.13) ∀i. Vk [jk]). k ∈ Vk [jk] [Cormen et al. where |V (i) | is the number of nodes in the graph for example i.4. ∀jk.j k i. (4. qi (w. this formulation requires that the upper bound on the empirical risk be zero.. A basic property of a spanning tree is that cutting any edge (j.j k + i. such as in perfect matching optimality in Ch. Then the following joint convex program in w and ν computes the max-margin parameters: min s. RS [hw ] = 0. Suppose that we can find a compact convex formulation of these conditions via a polynomial (in Li ) set of functions qi (w. νi ) ≤ 0 ⇔ w fi (y(i) ) ≥ w fi (y) + i (y). yjk ).t. (4.jk . νi ). yjk = 1. ∀y ∈ Y (i) . suppose yjk encodes whether edge (j. where j ∈ Vj [jk] and k ∈ Vk [jk]. 10. For a full graph. and minimizes the complexity . k ∈ Vk [jk]. (i) Then the optimality conditions are: (i) .jk for some basis functions f (xjk . with loss for each wrong edge given by w fi. If the basis functions are not sufficiently rich to ensure that each y(i) is optimal. we have |V (i) |3 ) constraints for each example i.1. 1 ||w||2 2 qi (w. the weight of (j. We discuss the conditions for optimality of perfect matchings in Ch. Essentially. k) in the tree creates two disconnected sets of nodes (Vj [jk]. jointly convex in w and auxiliary variables νi : ∃νi s. k) is in the tree and the score of the edge is given by w fi. 2001].t. but in other problems. k) is larger than (or equal to) the weight of any other edge (j . A spanning tree is optimal with respect to a set of edge weights if and only if for every edge (j. 10. k) in the tree connecting Vj [jk] and Vk [jk]. Note that this formulation does not allow for violations of the margin constraints (it has no slack variables ξi ).t. νi ) ≤ 0. j ∈ Vj [jk]. We also assume that the loss function decomposes into a sum of losses over the edges.jk ≥ w fi. In the spanning tree problem. Expressing the spanning tree optimality does not require additional variables νi . then Eq.1. MAX-MARGIN ESTIMATION 49 Consider the maximum weight spanning tree problem. k ) in the graph with j ∈ Vj [jk]. such auxiliary variables are needed.

1 shows schematically how approximating of the max subproblem reduces or extends the feasible space of w and ξ and leads to approximate solutions. If we continue adding constraints until there are no more violated ones. identifying these constraints is in general as difficult as solving the problem. we do not have to include all of them. If we can identify a small number of constraints that are critical to the solution. but we consider two general cases below. but a greedy approach of adding the most violated constraints often achieves good approximate solutions after adding a small (polynomial) number of constraints. Often. but can only upper or lower bound it. Fig.1 Constraint generation When neither compact maximization or optimality formulation is possible. and therefore are less justified. Only a subset of those constraints will be active at the optimal solution w.50 CHAPTER 4. where we have an exponential number of linear constraints. Consider Eq. the designer could add more basis functions f (x.2 Approximations: upper and lower bounds There are structured prediction tasks for which we might not be able to solve the estimation problem exactly. (4. 4. since that is the number of variables. The nature of these lower and upper bounds depends on the problem. Of course. but the maximization problem can be solved or approximated by a combinatorial algorithm.3). 4. y) that take into account additional information about x. one for each i and y ∈ Y (i) .2. In fact. 4. However. If the problem is infeasible. the resulting solution is optimal. not more than the number of parameters n plus the number of examples m can be active in general. STRUCTURED MAXIMUM MARGIN ESTIMATION of the hypothesis hw as measured by the norm of the weights. we can resort to constraint generation or cutting plane methods. One could also add slack variables for each example and each constraint that would allow violations of optimality conditions with some penalty. we cannot compute maxy∈Y (i) [w fi (y) + i (y)] exactly or explicitly. We assume that we have an algorithm that produces y = arg maxy∈Y (i) [w fi (y) + . these slack variables would not represent upper bounds on the loss as they are in the min-max formulation.

(4. If it is violated. At each iteration. The parabolic contours represent the value of the objective function and ‘+’.3) and Eq. terminating when the desired precision is reached. 4.14) ∀i. We maintain. The solid red line represents the constraints imposed by the assignments y ∈ Y(i) . The algorithm terminates when no constraints are violated. In Fig. i (y)]. w fi (y(i) ) + ξi ≥ w fi (y) + i (y). The algorithm is described in Fig. and be more stringent or relaxed in others. we solve the problem with a subset of constraints: min s. whereas the dashed and dotted lines represent approximate constraints.4. The approximate constraints may coincide with the exact constraints in some cases.2.1. (4. for each example i.2.14) is that Y (i) has been replaced by Y (i) . APPROXIMATIONS: UPPER AND LOWER BOUNDS 51 Upper-bound Exact Lower-bound + × Figure 4. The only difference between Eq. is violated.t. We then compute y = arg maxy∈Y (i) [w fi (y) + i (y)] for each i and check whether the constraint w fi (y(i) ) + ξi + ≥ w fi (y) + i (y). ∀y ∈ Y (i) . the lower-bound on the constraints provided by Y (i) ∪ y keeps tightening with each iteration. we set Y = Y (i) ∪ y. where (i) is a user defined precision parameter. 1 ||w||2 + C 2 ξi i (4. a small but growing set of assignments Y (i) ⊂ Y (i) . 4. ‘x’ and ‘o’ mark the different optima. We note that if the algorithm that produces y = arg maxy∈Y (i) [w fi (y)+ i (y)] is .1: Exact and approximate constraints on the max-margin quadratic program.

the approximation error of the solution we achieve might be much greater than . suboptimal. 1999]. 11. The number of constraints that must be added before the algorithm terminates depends on the precision and problem specific characteristics. Initialize: Y (i) = {}.2. Return w. 3. The relaxation usually defines a larger feasible space Y (i) ⊃ Y (i) over . if w fi (y(i) ) + ξi + ≤ w fi (y) + i (y). 2004] for a more in-depth discussion of cutting planes methods.2: A constraint generation algorithm. ∀ i. but very large. 1 ||w||2 + C 2 (i) ξi i w fi (y ) + ξi ≥ w fi (y) + i (y). multi-way cut in Ch. 7. while the maximization algorithm is comparatively fast.52 CHAPTER 4. graph-partitioning in Ch. Relaxations of such integer programs into LPs. QPs or SDPs often provide excellent approximation algorithms [Hochbaum. Compute y = arg maxy∈Y (i) [w fi (y) + i (y)]. Boyd & Vandenberghe. the maximization problem we are interested in may be very expensive or intractable. 4. 1. 1997. ∀i. we consider MAP inference in large tree-width Markov networks in Ch. For example. See [Bertsimas & Tsitsiklis. For each i. ∀y ∈ Y (i) . 8. then set Y (i) = Y (i) ∪ y and violation = 1 4.2 Constraint strengthening In many problems.t. Set violation = 0 and solve for w and ξ by optimizing min s. 2. Nemhauser & Wolsey. STRUCTURED MAXIMUM MARGIN ESTIMATION Input: precision parameter . Many such problems can be written as integer programs. Figure 4. 1997. This approach may also be computationally faster in providing a very good approximation in practice if the explicit convex programming formulation is polynomial in size. if violation = 1 goto 2.

the approximate MAP solution has higher value than any integer solution. the estimation algorithm is finding weights that are not necessarily optimal for an exact maximization algorithm. including the true assignment y(i) . 2004]. 1996. An inverse optimization problem is defined by an instance of an optimization problem maxy∈Y w f (y). and a target solution yt . the approximation is an overestimate of the constraints: max [w fi (y) + i (y)] ≥ max [w fi (y) + i (y)]. In such a case. where y ∈ Y (i) may correspond to a “fractional” assignment. In practice. Note that for every setting of the weights w that produces fractional solutions for the relaxation. (3. a solution to the MAP LP in Eq. By contrast. but (close to) optimal for the particular approximate maximization algorithm used. the margin has the standard interpretation as the difference between the score of y(i) and the MAP y (according to w). 4. Ahuja & Orlin. Fig.3. for weights w for which the MAP approximation is integer-valued. 2001. which make the . For example.1 shows how the over-estimate reduces the feasible space of w and ξ. 4. Zhang & Ma. As the objective includes a penalty for the slack variable.3 Related work Our max-margin formulation is related to a body of work called inverse combinatorial and convex optimization [Burton & Toint. In this case.4. a set of nominal weights w0 . minimizing the objective tends to drive the weights w away from the regions where the solutions to the approximation are fractional. In essence. 1992. The goal is to find the weights w closest to the nominal w0 in some norm. the approximate constraints are tightened because of the additional invalid assignments. intuitively.2) for an untriangulated network may not correspond to any valid assignment. y∈Y (i) y∈Y (i) Hence the constraint set is tightened with such invalid assignments. thereby driving up the corresponding slack ξi . we will show experimentally that such approximations often work very well. RELATED WORK 53 which the maximization is done. Heuberger.

||w − w0 ||p w f (yt ) ≥ w f (y). To map this problem (approximately) to our setting. The inverse reinforcement learning problem [Ng & Russell. then w = 0 is trivially the optimal solution. The study of inverse problems began with geophysical scientists (see [Tarantola. Abbeel & Ng. ∀y ∈ Y. but L2 norm is also used. the goals are very different than ours. a state to action mapping) that maximizes its expected reward. At every time step. The environment is modeled as a system that can be in one of a set of discrete states. 1987] for in-depth discussion of a wide range of applications). A full description of the problem is beyond our scope. Although there is a strong connection between inverse optimization problems and our formulations. The agent collects a reward at each step. note that a policy roughly corresponds to the labels . Most of the attention has been on L1 and L∞ norms. Provided educated guesses for the parameters w0 and the behavior of the system as a target. STRUCTURED MAXIMUM MARGIN ESTIMATION target solution optimal: min s. which is not appropriate in the learning setting. we are learning a parameterized objective function that depends on the input x and will generalize well in prediction on new instances. 2000. The solution w depends critically on the choice of nominal weights. A rational agent executes a policy (essentially. but we briefly describe the Markov decision process (MDP) model commonly used for sequential decision making problems where an agent interacts with its environment. In our framework. Modeling a complex physical system often involves a large number of parameters which scientists find hard or impossible to set correctly. the agent chooses an action from a discrete set of actions and the system transitions to a next state with a probability that depends on the current state and the action taken. Moreover.t. Note that if we set w0 = 0. the inverse optimization problem attempts to match the behavior while not perturbing the “guesstimate” too much.54 CHAPTER 4. The goal is to learn a reward function that will cause a rational agent to act similar to the observed behavior of an expert. we do not assume as given a nominal set of weights. which generally depends on the on the current and the next state and the action taken. 2004] is much closer to our setting.

the state sequence correspond to the input x and the reward for a state/action sequence is assumed to be w f (x. many other problems in which . (such as low-treewidth Markov networks. to learn a probabilistic model P (y | x) over bipartite matchings using maximum likelihood requires computing the normalizing partition function. certificate. Because our approach only relies using the maximum in the model for prediction. maximum margin estimation can be tractable when maximum likelihood is not. The estimated parameters are then approximate. The first formulation. This and related problems are formulated as a convex program in Ng and Russell [2000] and Abbeel and Ng [2004]. CONCLUSION 55 y. Similar results hold for non-bipartite matchings and min-cuts.4. 4. 1979. but produce accurate approximate prediction models in practice. Our framework applies to a wide range of prediction problems that we explore in the rest of the thesis. and does not require a normalized distribution P (y | x) over all outputs. min-max. The goal is to learn w from a set of state/action sequences (x(i) .4. By contrast. y). including Markov networks. context free grammars. which is #P-complete [Valiant. and many combinatorial structures such as matchings and graph-cuts. The second one. only requires expressing optimality of a given assignment according to the model. maximum margin estimation can be formulated as a compact QP with linear constraints. When the prediction problem is intractable or very expensive to solve exactly.4 Conclusion In this chapter. we presented two formulations of structured max-margin estimation that define a compact convex optimization problem. For example. Garey & Johnson. y) for some basis functions w f (x. The estimation problem is tractable and exact whenever the prediction problem can be formulated as a compact convex optimization problem or a polynomial time combinatorial algorithm with compact convex optimality conditions. y(i) ) of the expert such that the maximizing the expected reward according to the system model makes the agent imitate the expert. In models that are tractable for both maximum likelihood and maximum margin. context free grammars. relies on the ability to express inference in the model as a compact convex optimization problem. 1979]. we resort to approximations that only provide upper/lower bounds on the predictions.

the solutions to the estimation are relatively sparse in the dual space (as in SVMs). Altun et al. which makes the use of kernels much more efficient. 2004]. . The range of applications of these principles is very broad and leads to estimation problems with very interesting structure in each particular problem. our approach has an additional advantage. The forthcoming formulations in the thesis follow the principles laid out in this chapter. Maximum likelihood estimation with kernels results in models that are generally non-sparse and require pruning or greedy support vector selection methods [Lafferty et al.. 2004. Because of the hinge-loss.. STRUCTURED MAXIMUM MARGIN ESTIMATION inference is solvable by dynamic programming).56 CHAPTER 4. from Markov networks and context-free grammars to graph cuts and perfect matchings.

Part II Markov networks 57 .

2002]. a Markov network for the hypertext domain would include a node for each webpage. the labels (letters) of adjacent inputs (images) are highly correlated. 1998.. because one links to the other). and many other fields. and an edge between any pair of webpages whose labels are directly correlated (e. In image processing. 1998]). Markov networks compactly represent complex joint distributions of the label variables by modeling their local interactions. location and function of proteins that interact are often highly correlated [Vazquez et al. In hypertext or bibliographic classification. Taskar et al. neighboring pixels exhibit spatial label coherence in denoising. 1999a]. For example.. Sequential prediction problems arise in natural language processing (part-of-speech tagging.g. segmentation and stereo correspondence [Besag. whose nodes represent the different object labels.Chapter 5 Markov networks Markov networks are extensively used to model complex sequential. encoding its label. compuu tational biology (gene finding. and relational interactions in prediction problems arising in many fields. These problems involve labeling a set of related objects that exhibit local consistency. We address the problem of max-margin estimation the parameters of Markov networks 58 . Such models are encoded by a graph. In sequential labeling problems (such as handwriting recognition).. Boykov et al. 1986. protein structure prediction.... labels of linked and co-cited documents tend to be similar [Chakrabarti et al. information extraction [Manning & Sch¨ tze. and whose edges represent and quantify direct dependencies between them. 1999]). 2003]. sequence alignment [Durbin et al. In proteomic analysis. spatial. speech recognition.

3. y). this mapping depends on the segmentation algorithm that determines how many letters the sample image contains and splits the image into individual images for each letter. MAXIMUM LIKELIHOOD ESTIMATION 59 for such structured classification problems.. for example. 5. y): log Pw (y | x) = w f (x. 2001]). The objective is to minimize the training log-loss with an additional regularization term.1 Maximum likelihood estimation The regularized maximum likelihood approach of learning the weights w of a Markov network is similar to logistic regression we described in Sec. In handwriting recognition.2.1. The log-linear representation we have described in Sec. It also depends on the basis functions we use to model the dependencies of the problem.. usually the squared-norm of the weights w [Lafferty et al. We analyze the theoretical generalization properties of max-margin estimation and derive a novel margin-based bound for structured classification. Before we present the maximum margin estimation. where Zw (x) = y exp{w f (x.1 is defined via a vector of n basis functions f (x. i i . 2. E} where the nodes V map to the variables in y. y) − log Zw (x). y(i) )}m . we review the standard maximum likelihood method. drawn from a fixed disi=1 tribution D over X × Y. 3. We are given a labeled training sample S = {(x(i) . y)} and w ∈ IRn . which represent P (y | x) instead of generative models P (x. We abbreviate G(x(i) ) as G (i) below. We focus on conditional Markov networks (or CRFs [Lafferty et al. Note that the topology and size of the graph G (i) .5. 2001]: 1 ||w||2 − C 2 1 log Pw (y(i) | x(i) ) = ||w||2 + C 2 log Zw (x(i) ) − w fi (y(i) ). We show a compact convex formulation that seamlessly integrates kernels with graphical models. first-order Markov chain or a higher-order models. We assume the structure of the network is given: we have a mapping from an input x to the corresponding Markov network graph G(x) = {V. the training sequences might have different lengths. might be different for each example i. For instance.

1999. To compute the expectations.60 CHAPTER 5. Pw (yc∪c | xi ) as well as covariances of all basis functions. which is quadratic in the number of cliques and the number of functions. we can use inference in the Markov network to calculate marginals Pw (yc | x(i) ) for each clique c in the network Sec.w δfi (y)δfi (y) . Let δfi (y) = fi (y) − Ei. MARKOV NETWORKS where fi (y) = f (x(i) . 2004].3. y). y). even with millions of parameters w.w [fi (y)] − fi (y(i) ) = w − C i y∈Y fi (y)Pw (y Ei. y(i) ) − f (x(i) . so we have an unconstrained convex optimization problem. Pinto et al.w [fi (y)]. require the second derivatives as well as the gradient. as before. The Hessian of the objective depends on the covariances of the basis functions: I +C i Ei.w [fi (y)] = | x(i) ) is the expectation under the conditional dis- tribution Pw (y | x(i) ) and ∆fi (y) = f (x(i) . 3. A standard approach is to use an approximate second order method that does not need to compute the Hessian. The gradient with respect to w is given by: w+C i Ei. Conjugate Gradients or L-BFGS methods have been shown to work very well on large estimation problems [Sha & Pereira. Boyd & Vandenberghe. where Ei.. since we need to calculate joint marginals of every pair of cliques c and c . . Since the basis functions decompose over the cliques of the network. but uses only the gradient information [Nocedal & Wright. the expectation decomposes as well: Ei. 2003.w [fi (y)] = (i) c∈C (i) yc ∈Yc fi. where I is a n × n identity matrix. 2003].w [∆fi (y)]. Computing the Hessian is more expensive than the gradient. such as Newton’s method.c (yc )Pw (yc | x(i) ).2. Second order methods for solving unconstrained convex optimization problems. This objective function is convex in the parameters w.

Assumption 5. ∀c ∈ C (i) .yc µi. 3. but the important difference is the loss function i . ∀i. The simplest loss is the 0/1 loss i (y) ≡ 1 (i) = y).2. I(y (i) With this assumption. in structured problems. like the basis functions.c (yc )] (5. we assume that the loss. yc µi. A natural choice that we use in our experiments is the Hamming distance: H (x(i) . where we are predicting multiple labels. s ⊂ c. s.2) µi. In general.s (ys ) = yc ∼ys µi. In fact I(y this loss for sequence models was used by Collins [2001] and Altun et al.3 how to express maxy w fi (y) as an LP. y(i) . ∀ys . ∀i.3. but may depend on the number of labels and type of labels predicted incorrectly or perhaps the number of cliques of labels predicted incorrectly. 3. the loss is often not just the simple 0/1 loss. ∀yc . ∀c ∈ C (i) . µi.c (yc ) = 1.c (yc ) + i. w fi (y(i) ) + ξi ≥ max [w fi (y) + i (y)]. c ∈ C (i) . [2003]. we begin with the min-max formulation from Sec.t. .c (yc ). However.2 Maximum margin estimation For maximum-margin estimation.5.1: min 1 ||w||2 + C 2 ξi i y (5.t.3. We will focus on decomposable loss functions below. y) = v∈V (i) 1 v = yv ). 4.2.3: max c. we can express this inference problem for a triangulated graph as a linear program for each example i as in Sec. yc ) = c c∈C(G (i) ) i.c (yc ).c (yc ) ≥ 0.1) s.c (yc )[w fi.1 The loss function i (y) is decomposable: i (y) = c∈C(G (i) ) (i) (x(i) . MAXIMUM MARGIN ESTIMATION 61 5. We know from Sec. decomposes over the cliques of labels. yc . ∀s.

(5. For readability.c . λ and m).s.62 CHAPTER 5.c (yc ) + i. ys ∼yc In order to gain some intuition about this formulation. then the distribution that maximizes the objective puts all its weight on that assignment. we have: min s. ys ∼yc In this dual. s⊂c. we make a change of variables from λi. while mi. s.c + ∀c ∈ C (i) . ∀i.c (yc ) = s⊂c. ∀c ∈ C (i) . we also introduce variables that capture the effect of all the agreement variables m: Mi. MARKOV NETWORKS where C (i) = C(G (i) ) are the cliques of the Markov network for example i.c = w fi.4) λi.s.s. ∀i. The reason for naming the new variables using the letter ξ will be clear in the following.c + s⊃c mi.c (yc ) + ξi. λi.c.c (yc ) − mi. ∀i. ∀c ∈ C (i) . If the most likely assignment is unique.c.c to ξi. the λi. the constraints ensure that the µi ’s form a proper distribution. (If the arg max is not unique.c (yc ) + i.2) is given by: min c λi. As we showed before.c. any convex combination of the assignments is a valid solution). ∀c ∈ C (i) . The dual of Eq.3) mi.c (yc ). ξ. ∀yc .c : (i) λi. s⊂c.s (ys ) ≥ w fi. ys ∼yc mi.s (yc ) variables correspond to the agreement constraints in the primal in Eq.c (yc ) − s⊃c (5.c mi. ∀yc .c w fi (y(i) ) + ξi ≥ λi.c (yc ).2). . i.c. ∀yc . (5.1) for each example i and maximizing jointly over all the variables (w.s (ys ) ≥ w fi. Plugging the dual into Eq.c (yc ).t. (5.c .t.c variables correspond to the normalization constraints. 1 ||w||2 + C 2 ξi i (5.s (ys ) − s⊃c mi.

c (yc ) + ξi.1).s variables. MAXIMUM MARGIN ESTIMATION 63 With these new variables.c (yc )].c (yc ). (5. (i) w fi. for each clique: min s.c yc i. we can write this in a form that resembles our original formulation Eq.c (yc ) = s⊂c. ∀i. ∀i. since the slack variable ξi only appears only in the and the objective minimizes Cξi .c. we have: min s.c (yc ) ξi i (5.c (yc ) + i.s (ys ) variables correspond to a certain kind of messages between cliques that distribute “credit” to cliques to fulfill this margin requirement from other cliques which have sufficient margin.s (ys ) − s⊃c mi. 1 ||w||2 + C 2 ξi. ∀yc . ∀i.c.c this set of variables: min s. but defined at a local level. Note that without Mi. ∀i.c.5) ξi.c. ∀c ∈ C (i) .c (yc ) + Mi.c i. ∀yc . ∀c ∈ C (i) .c (yc ) = s⊂c.c (yc ) (5.c (yc ) + Mi.c (yc ).c.c (yc ) + ξi. 1 ||w||2 + C 2 ξi ≥ c (i) w fi. ∀i.6) + Mi.c i.c (yc ).c and mi.c ≥ w fi.7) + Mi. Note that ξi = constraint ξi ≥ c ξi.s (ys ) − s⊃c mi. ∀c ∈ C (i) . ∀i. . (i) w fi.t.5.s.c at the optimum. 1 ||w||2 + C 2 ξi.c (yc ). we essentially treat each clique as an independent classification problem: for each clique we have a hinge upper-bound on the local loss.2. ∀c ∈ C (i) . ∀c ∈ C (i) . ys ∼yc mi.c (yc ) = s⊂c. ys ∼yc mi.t. ∀yc . Finally.s (ys ) − s⊃c mi.c .s.t. Mi.c i. ∀c ∈ C (i) .c ≥ max [w fi.c ≥ w fi. ∀i. + Mi. ∀yc .s. or a margin requirement. The mi. Hence we can simply eliminate c ξi. ∀yc .c (yc ) + ξi.c (yc ) (5.c (yc ). ys ∼yc mi.

Suppose for the sake of this example that our training data consists of only one training sample.45 (y5 ).23 (y2 ). For concreteness. 5.1. I(y w f3 (y3 ) + ξ3 ≥ w f3 (y3 ) + 1 3 = y3 ) − m3. ∀y2 . MARKOV NETWORKS Figure 5.23 (y3 ) − m3.34 (y4 ) − m4. I(y w f4 (y4 ) + ξ4 ≥ w f4 (y4 ) + 1 4 = y4 ) − m4. ∀y1 . The set of cliques consists of the five nodes and the four edges. I(y (∗) (∗) (∗) (∗) (∗) (∗) (∗) (∗) (∗) (∗) . below we use the Hamming j (yj ) (i) . The constraints associated with the node cliques in this sequence are: w f1 (y1 ) + ξ1 ≥ w f1 (y1 ) + 1 1 = y1 ) − m1.45 (y4 ). Also shown are the corresponding local slack variables ξ for each clique and messages m between cliques. For brevity of notion in this example. we drop the dependence on the sample index i in the indexing of the variables (we also used yj loss H (∗) instead of yj below).64 CHAPTER 5. As an example. ∀y4 .12 (y2 ) − m2. I(y w f5 (y5 ) + ξ5 ≥ w f5 (y5 ) + 1 5 = y5 ) − m5.1: First-order chain shown as a set of cliques (nodes and edges). ∀y3 .12 (y1 ). I(y w f2 (y2 ) + ξ2 ≥ w f2 (y2 ) + 1 2 = y2 ) − m2. ∀y5 .34 (y3 ). consider the first-order Markov chain in Fig. The figure shows the local slack variables ξ and messages m between cliques for this sample. which decomposes into local terms = 1 j = yj ) for each node and is I(y (∗) zero for the edges.

y(i) ) − f (x(i) . y2 ) + m1. y5 . y4 ) + ξ34 ≥ w f34 (y3 . w f34 (y3 .3. we develop an alternative dual derivation that provides a very interesting interpretation of the problem and is a departure for special-purpose algorithms we develop. M3 N DUAL AND KERNELS 65 The edge constraints are: w f12 (y1 . 2 αi (y)∆fi (y) i. y3 ) + ξ23 ≥ w f23 (y2 .34 (y4 ). ∀y4 .12 (y1 ) + m2. y5 ) + ξ45 ≥ w f45 (y4 . y4 .45 (y4 ) + m5. y. so that they sum . ∀y3 .y 1 αi (y) i (y) − 2 αi (y) = C.3 M3N dual and kernels In the previous section. 1 ||w||2 + C 2 ξi i (5.3): min s. The dual is given by: max i.t.23 (y2 ) + m3. y αi (y) ≥ 0. (4. y.8) w ∆fi (y) ≥ i (y) − ξi . y). ∀i. ∀i. w f23 (y2 . y2 .5. We begin with the formulation as in Eq. y3 . We make two small transformations to the dual that do not change the problem: we normalize α’s by C (by letting αi (y) = Cαi (y)).t. ∀y1 . the exponential number of αi (y) variables correspond to the exponential number of constraints in the primal. w f45 (y4 . ∀y2 . y5 ) + m4.23 (y3 ). y2 ) + ξ12 ≥ w f12 (y1 .9) s.45 (y5 ). y4 ) + m3. we showed a derivation of a compact formulation based on LP inference. In this section. In the dual. (∗) (∗) (∗) (∗) (∗) (∗) (∗) (∗) 5.34 (y3 ) + m4. where ∆fi (y) ≡ f (x(i) . ∀i. y3 ) + m2.12 (y2 ).y (5.

The dual objective is a function of expectations of i (y) and ∆fi (y) with respect to αi (y).10) can be interpreted as a kind of distribution over y. y αi (y) ≥ 0. Note that the number of µi. the first term in the objective function (fixing a particular i): αi (y) i (y) = y y αi (y) c i.c (yc ) and ∆fi (y) = c ∆fi.10) in terms of these marginal dual variables. y αi (y) ≥ 0.66 CHAPTER 5. We define the marginal dual variables as follows: µi.c (yc ) variables is small (polynomial) compared to the number of αi (y) variables (exponential) if the size of the largest clique is constant with respect to the size of the network.c (yc ) y ∼yc αi (y ) = c. y. Now we can reformulate our entire QP (5. Since i (y) = c i. ∀yc .t.c (yc ) i. for example. ∀y.y ∗ αi (y)∆fi (y).y 1 αi (y) i (y) − C 2 αi (y) = 1.10) s. since they lie in the simplex αi (y) = 1. This dual distribution does not represent the probability that the model assigns to an instantiation. 2 αi (y)∆fi (y) i. ∀c ∈ C (i) . The resulting dual is given by: max i. the solution to the dual α gives the solution to the primal as a weighted combination: w∗ = C i. . Consider.c (yc ) = c. MARKOV NETWORKS to 1 and divide the objective by C.11) where y ∼ yc denotes whether the partial assignment yc is consistent with the full assignment y . As in multi-class SVMs.c (yc ). Our main insight is that the variables αi (y) in the dual formulation Eq.c (yc ) decompose over the cliques of the Markov network. ∀i.y (5. (5.c (yc ) = y ∼yc αi (y ). (5. ∀i.yc µi. we only need clique marginals of the distribution αi (y) to compute their expectations. but the importance of the constraint associated with the instantiation to the solution.yc i. ∀i.

(5. µi.5).s (ys ) = yc ∼ys µi. s ⊂ c. the α distribution is not unique.c. but do not make a commitment about what it is. D[M]. Q(α) = Q (M(α)). ∀c ∈ C (i) . c ∈ C (i) . In general.c. the following structured dual QP has the same primal solution (w∗ ) as the original exponential dual QP in Eq. the set of legal marginals? Characterizing this set (also known as marginal polytope) compactly will allow us to work in the space of µ’s: max Q(α) ⇔ max Q (µ). i. R[M]. ∀c ∈ C (i) . The domain of this operator. which is exactly what the constraints in the LP for MAP inference enforce (see discussion of Lemma 3.c (yc ) i. Therefore. that is.c (yc )∆fi.c (yc ) − C 2 2 µi. ∀i.c (yc ).c (yc ).c (yc ) i.yc ∆fi. ∀ys . The objective of .c (yc ). but there is a continuum of distributions consistent with a set of marginals. ∀yc . M3 N DUAL AND KERNELS 67 The decomposition of the second term in the objective is analogous.c (yc )∆fi.12) s.yc 1 µi.5. ∀s. µi.c.c (yc ). where M denotes the marginalization operator defined by Eq. µ∈R[M] α∈D[M] Hence we must ensure that µi corresponds to some distribution αi .c (yc ) = 1. What is its range. Note that it only depends on αi (y) through its marginals µi.yc µi. (5.11) .3.c In this structured dual. when all G (i) are triangulated.c (yc ) y ∼yc αi (y ) = c. we only enforce that there exists an αi consistent with µi . Let us denote the the objective of Eq.10) as Q(α).10): max i.t.3.c (yc ) ≥ 0. ∀i. yc µi. is the product of simplices for all the m examples.yc (5. ∀i.yc µ∗ (yc )∆fi. (5. The solution to the structured dual µ∗ gives us the primal solution: w∗ = C i. αi (y)∆fi (y) = y c.

c (yc ) ∆fj. xc ). e. by convention 0/0 ≡ 0. 1999]. (5.c (yc ) . 6. c ¯ ¯ ¯ ¯ ¯ c c (j) (j) (j) Hence. We denote the set of separators as S (i) . In Sec.10) does not distinguish between these distributions. each log-potential in the network “remembers” only a small proportion of the relevant training data inputs. yc ) − f (x(i) . As in SVMs. is unique for a triangulated model and can be computed using the junction tree T (i) for the network [Cowell et al.s (ys ) (5. yc ) f (xc . We usually extend the kernel over the input space to the joint input and output space by simply defining f (xc .12) can be expressed in terms of dot products between local basis functions (i) ∆fi. the solutions to the max-margin QP are typically sparse in the µ variables.g. Specifically. I(y ¯ ¯ ¯ ¯ Of course. k(xc . since it only depends on their marginals. we can locally kernelize our models and solve Eq.13) Kernels Note that the solution is a weighted combination of local basis functions and the objective of Eq. xc ). s∈S (i) µi.68 CHAPTER 5. Note that each separator s and complement of a separator c \ s is also a clique of the original graph. yc ) ≡ 1 c = yc )k(xc . MARKOV NETWORKS the QP Eq.¯(yc ) = [f (x(i) . c ) in the tree T (i) is a set of variables called the separator s = c ∩ c . other definitions are possible and may be useful when the assignments in each clique yc have interesting structure. associated with each edge (c..12) efficiently.2 we experiment with several kernels for the handwriting example. yc ) − f (xc . c∈T (i) µi. Kernels are typically defined on the input. (i) (j) . Hence. yc )] [f (xc . (5. yc )]. however. In our handwriting example. (5. we use a ¯ polynomial kernel on the pixel values for the node cliques. The maximum-entropy distribution αi consistent with a set of marginals µi . since it is a subclique of a larger clique. Now we can define the maximum-entropy distribution αi (y) as follows: αi (y) = Again.

y3 . we must address the problem by triangulating the graph. adding fill-in edges to ensure triangulation. UNTRIANGULATED MODELS 69 Figure 5. Y3 and Y1 . y3 . y1 ∀y2 . Y4 and the corresponding marginals. we can triangulate the graph by adding an arc Y1 —Y3 . y2 . y3 ). that is.5. y2 . µ123 (y1 . Y2 . y3 . y3 . y2 . This will introduce new cliques Y1 . We can then use this new graph to produce the constraints on the marginals: µ123 (y1 . y3 µ134 (y1 .4.2. ∀y3 . y4 .4 Untriangulated models If the underlying Markov net is not chordal. µ123 (y1 . y3 ). y2 ). y2 . 5. ∀y1 . y4 ). Y3 . y3 ) and µ134 (y1 . they do not add any new basis functions nor change the objective function. y4 ) = µ34 (y3 .2: Diamond Markov network (added triangulation edge is dashed and three-node marginals are in dashed rectangles). y1 µ134 (y1 . . y4 ). y3 The new marginal variables appear only in the constraints. y3 ) = µ23 (y2 . For example. 5. y3 . y3 ) = µ12 (y1 . y4 ) = µ13 (y1 . ∀y1 . if our graph is a 4-cycle Y1 —Y2 —Y3 —Y4 —Y1 as in Fig.

which enforces only local consistency of marginals. for example 2D grids often used in image analysis. Solving the problem as C → ∞ will find this solution even though it may not be optimal (in terms of the norm of the w) using exact inference. we show a generalization bound for the task of structured classification that allows us to relate the error rate on the training set to the generalization error. 5. 8 demonstrate.5 Generalization bound In this section. optimizes our objective only over a relaxation of the marginal polytope. Such a formulation. To the best . which tend to improve the approximation of the marginal polytope. we can add positive semidefinite constraints on the marginals µ used by Wainwright and Jordan [2003]. (5. However.12) defined by an untriangulated graph.70 CHAPTER 5. even sparsely connected networks.6) or the structured dual Eq. The approximate QP does not guarantee that the learned model using exact inference minimizes the true objective: (upper-bound on) empirical risk plus regularization. Unfortunately. However. consider the separable case. the learned parameters produce very accurate approximate models in practice. But do we really need these optimal parameters if we cannot perform exact inference? A more useful goal is to make sure that training error is minimized using the approximate inference procedure via the untriangulated LP. Although this and other more complex relaxations are a very interesting area of future development. where 100% accuracy is achievable on the training data by some parameter setting w such that approximate inference (using the untriangulated LP) produces integral solutions. we can still solve the QP in the structured primal Eq. We conjecture that the parameters learned by the approximate QP in fact do that to some degree. MARKOV NETWORKS In general. Note that we could also strengthen the untriangulated relaxation without introducing an exponential number of constraints. the number of constraints introduced is exponential in the number of variables in the new cliques — the tree-width of the graph. the formulation trades off fractionality of the untriangulated LP solutions with complexity of the weights ||w||2 . they are often much more expensive. For example. For instance. as experiments in Ch. have large tree-width. (5. For C in intermediate range.

Consider an upper bound on the above loss: L(w. An appropriate error function is the average per-label loss L(w. x. Our goal in structured classification is often to minimize the number of misclassified labels. y where L is the number of label variables in y. we need to define several other factors it depends upon. Let Nc be the maximum number of cliques in G(x). y) = max 1 H (y. Consider a first-order sequence . y ).5. Otherwise. and Rc be an upper-bound on the 2-norm of clique basis functions. This upper bound is even more adversarial: it is tight if y = arg maxy [w f (x. Note that the loss we minimize in the maxmargin formulation is very closely related (although not identical to) this upper bound. arg max w f (x.To state the bound. x. This upper bound is tight if y = arg maxy w f (x. it picks from all y which are better when helped by γ H (y. or the Hamming distance between y and h(x). y ). y ). it is adversarial: it picks from all y which are better (w f (y) ≤ w f (y )). otherwise. q be the maximum number of cliques that have a variable in common. one that maximizes the Hamming distance from y. y) ≤ L(w. as we detail below. We can now define a γ-margin per-label loss: L(w.5. y) = max 1 f (y ) L H y : w f (y)≤w (y. As in other generalization bounds for marginbased classifiers. Vc be the maximum number of values in a clique |Yc |. GENERALIZATION BOUND 71 of our knowledge. x. plus a term that grows inversely with the margin. y) ≤ L(w. Our analysis of Hamming loss allows to prove a significantly stronger result than previous bounds for the 0/1 loss. y ) + H (y. x.y ) L H y : w f (y)≤w f (y )+γ (y. y ). this bound is the first to deal with structured error. y) = 1 L H (y. one that maximizes the Hamming distance from y. such as the Hamming distance. We can now prove that the generalization accuracy of any hypothesis w is bounded by its empirical γ-margin per-label loss. x. we relate the generalization error to the margin of the classifier. y )). x. y) ≤ Lγ (w. y )].

This is a significant gain over the previous result of Collins [2001] for 0/1 loss. MARKOV NETWORKS model as an example. and V the number of values a variable takes. yielding significant improvements in the bound. Theorem 5. x. 2001]. x. Finally. note that if of the number of labels L. However we propose a novel method for covering the space of structured prediction models by using a cover of the individual clique basis function differences ∆fi. Then Nc = 2L − 1 since we have L node cliques and L − 1 edge cliques.c (yc ). and depends on the joint 2-norm of all of the basis functions for an example (which is ∼ Nc Rc ). node clique. the expected per-label loss is bounded by: ED [L(w. until now. y)] at high margin γ quantifies the confidence of the prediction model. in OCR. The first term upper bounds the training error of w.1 for the proof details and the exact value of the constant K. then our bound is independent . Proof: See Appendix A. L m = O(1) (for example. and current-next edge clique. and q = 3 since nodes in the middle of the sequence participate in 3 cliques: previous-current edge clique. and any δ > 0. y)] ≤ ES [Lγ (w. with probability at least 1 − δ. The proof uses a covering number argument analogous to previous results in SVMs [Zhang. the result provides a bound to the generalization error that trades off the effective complexity of the hypothesis space with the training error.72 CHAPTER 5.5. with L as the maximum length. there exists a constant K such that for any γ > 0 per-label margin.1 For the family of hypotheses parameterized by w. This new type of cover is polynomial in the number of cliques. Specifically. x. an open problem for margin-based sequence classification [Collins. which corresponds to the complexity of the classifier (normalized by the margin level). for sequences. Thus. 2002]. if the number of instances is at least a constant times the length of a word). and m > 1 samples. Low loss ES [Lγ (w. our bound has a logarithmic dependence on the number of cliques (ln Nc ) and depends only on the 2-norm of the basis functions per-clique (Rc ). Vc = V 2 because of the edge cliques. y)] + 2 K Rc ||w||2 q 2 1 [ln m + ln Nc + ln Vc ] + ln m γ2 δ . The second term depends on ||w||/γ. Such a result was. which has linear dependence (inside the square root) on the number of nodes (L).

we achieve a much tighter analysis. computational biology. [2001] has inspired many applications in natural language. computer vision and relational modeling [Sha .6 Related work The application of margin-based estimation methods to parsing and sequence modeling was pioneered by Collins [2001] using the Voted-Perceptron algorithm [Freund & Schapire. Remarkably. In a follow-up paper.5.2. which is prohibitively expensive for large number of examples and label variables. the bound does not explicitly depend on the length of the sequence. By considering the more natural Hamming loss. However. The work of Guestrin et al. [2003] have applied the exponential-size formulation with constraint generation we described in Sec. The work of Lafferty et al. however. the corresponding problem needs to be resolved (or at least approximately resolved) after each additional constraint is added.1 to problems natural language processing. has linear dependence (inside the square root) on the number of nodes (L). Describing the exact setting is beyond our scope. Tsochantaridis et al. He provides generalization guarantees (for 0/1 loss) that hold for separable case and depend on the number of mistakes the perceptron makes before convergence. but it suffices to say that our original decomposition of the max-margin QP was inspired by the proposed technique to transform an exponential set of constraints into a polynomial one using a triangulated graph. RELATED WORK 73 5. 4. It also depends on the joint 2-norm of all of the basis functions for an example (which is ∼ Nc Rc ). [2003] presents LP decompositions based on graphical model structure for the value function approximation problem in factored MDPs (Markov decision processes with structure). the number of constraints in many important cases is several orders higher (in L) than in the the approach we present. 1998]. Collins [2004] also suggested an SVM-like formulation (with exponentially many constraints) and a constraint generation method for solving it. [2004] show that only a polynomial number of constraints are needed to be generated to guarantee a fixed level of precision of the solution. His generalization bound (for 0/1 loss) based on the SVM-like margin.6. although undoubtedly the number of mistakes does. There has been a recent explosion of work in maximum conditional likelihood estimation of Markov networks. Altun et al. In addition.

2003. Sutton et al. the solutions are non-sparse and the proposed algorithms are forced to use greedy selection of support vectors or heuristic pruning methods. Kumar & Hebert. 2004. 2004. compact. Lafferty et al. Our generalization bound significantly tightens previous results of Collins [Collins.7 Conclusion We use graph decomposition to derive an exact. Taskar et al.. 2004]. 2003b]. 5. However.. 2002. The seamless integration of kernels with graphical models allows us to create very rich models that leverage the immense amount of research in kernel design and graphical model decompositions. maximum conditional likelihood estimation for Markov networks can also be kernelized [Altun et al. we present an efficient algorithm that exploits graphical model inference and show experiments on a large handwriting recognition task that utilize the powerful representational capability of kernels. convex max-margin formulation for Markov networks with sequence and other low-treewidth structure. 2001] and suggests possibilities for analyzing per-label generalization properties of graphical models. 2003.. As in the case of logistic regression. .. Taskar et al.74 CHAPTER 5. We also use approximate graph decomposition to derive a compact approximate formulation for Markov networks in which inference is intractable.. In the next chapter. 2003. Pinto et al. MARKOV NETWORKS & Pereira. Our formulation avoids the exponential blow-up in the number of constraints in the max-margin QP that plagued previous approaches. We provide theoretical guarantees on the average per-label generalization error of our models in terms of the training set margin..

for standard QP solvers. αi (y)∆fi (y) i. which allows us to perform 75 . We show that our models significantly outperform other approaches by incorporating high-dimensional decision boundaries of polynomial kernels over character images while capturing correlations between consecutive characters. (5. s.y 1 αi (y) i (y) − C 2 αi (y) = 1. y αi (y) ≥ 0.y s. we use a coordinate dual ascent method analogous to the sequential minimal optimization (SMO) used for SVMs [Platt.Chapter 6 M3N algorithms and experiments Although the number of variables and constraints in the structured dual in Eq. ∀i. 2004] provide sufficient and necessary criteria for optimality of a dual solution α. Instead.1 Solving the M3N QP Let us begin by considering the primal and dual QPs for multi-class SVMs: 2 1 min ||w||2 + C 2 ξi i max i.t. As we describe below. 1999.12) is polynomial in the size of the data. ∀i. y. y. these conditions have certain locality with respect to each example i. unfortunately. w ∆fi (y) ≥ (y) − ξi . ∀i. The KKT conditions [Bertsekas. the problem is often too large even for small training sets. 1999].t. We apply our M3 N framework and structured SMO algorithm to a handwriting recognition task. Boyd & Vandenberghe. 6.

(KKT1) αi (y) > 0 ⇒ vi (y) ≥ vi (y). are optimal if they satisfy the following two types of constraints: αi (y) = 0 ⇒ w ∆fi (y) > i (y) − ξi . We can express these conditions as αi (y) = 0 ⇒ w fi (y) + i (y) < max [w fi (y ) + i (y )]. (6. (KKT2). we have αi (y) = 0 ⇒ vi (y) < vi (y). Con- versely. however. we will enforce KKT conditions up to a given tolerance 0 < αi (y) = 0 ⇒ vi (y) ≤ vi (y) + . αi (y) > 0 ⇒ vi (y) ≥ vi (y) − . (KKT1) (KKT2) αi (y) > 0 ⇒ w fi (y) + i (y) = max [w fi (y ) + i (y )]. y To simplify the notation. 1. αi (y) > 0 ⇒ w ∆fi (y) = i (y) − ξi . A feasible dual solution α and a primal solution defined by: w = C i. This allows us to perform dual block-coordinate ascent where a block corresponds to . we define vi (y) = w fi (y) + i (y).y y αi (y)∆fi (y) y (6. Note that the normalization constraints on the dual variables α are local to each example i.2) vi (y) = max [w fi (y ) + i (y )]. αi (y) can be non-zero only if vi (y) is at most smaller than the all others. αi (y) can be zero only if vi (y) is at most larger than the all others. With these definitions. Essentially.76 CHAPTER 6. M3 N ALGORITHMS AND EXPERIMENTS the search for optimal α by repeatedly considering one example at a time.1) ξi = max [ i (y) − w∆fi (y)] = max [ i (y) + w fi (y)] − w fi (y (i) ). y =y In practice. y (KKT1) (KKT2).

4. Only the second part of the objective Q(αi . For each i. Set violation = 0. y. Find feasible αi such that Q(αi . the algorithm will converge to the global maximum in a finite number of steps (within the precision).1).6.1 is essentially coordinate ascent on blocks αi . This allows us to focus on efficient updates to a single block of αi at a time. α−i ) matters for optimizing with respect to αi . Initialize: αi (y) = 1 = y (i) ). where α−i denotes all dual αk variables for k other than i. α−i ) in terms of λ and α: αj (y) j (y) + j. α−i ) > Q(αi . 6.1: Block-coordinate dual ascent. I(y 2. the objective function can be split into two terms: Q(α) = Q(α−i ) + Q(αi . 3. The algorithm starts with a feasible dual solution α and improves the objective block-wise until all KKT conditions are satisfied. the vector of dual variables αi for a single example i. Set violation = 1. (6. Figure 6. maintaining the feasibility of the dual. When optimizing with respect to a single block i. . The general form of block-coordinate ascent algorithm as shown in Fig.1. 7. Note that y λ(y) = 0 and αi (y) + λ(y) ≥ 0 so that αi is feasible. As long as the local ascent step over αi is guaranteed to improve the objective when KKT conditions are violated. ∀ i. Checking the constraints requires computing w and ξ from α according to Eq. Let αi (y) = αi (y) + λ(y). 5. α−i ). α−i ) and set αi = αi .y αj (y)∆fj (y) . If violation = 1 goto 2. SOLVING THE M3 N QP 77 1. We can write the objective Q(α−i ) + Q(αi . 6. If αi violates (KKT1) or (KKT2).y y 1 λ(y) i (y) − C 2 2 λ(y)∆fi (y) + y j.

1. All that is required is an ascent step. We address a strategy for selecting the two variables in the next section. we get: 1 − C 2 2 λ(y) i (y) − w y y λ(y)∆fi (y) λ(y)∆fi (y) y . not a full optimization.y αj (y)∆fj (y). (6. so we must change at least two variables in order to respect the normalization constraint (by moving weight from one dual variable to another).3) . M3 N ALGORITHMS AND EXPERIMENTS By dropping all terms that do not involve λ. but for now assume we have picked λ(y ) and λ(y ). ∀y. In our case. αi (y ) − δ ≥ 0. 1 [vi (y ) − vi (y )]δ − C||fi (y ) − fi (y )||2 δ 2 2 αi (y ) + δ ≥ 0.78 CHAPTER 6. and making the substitution w=C j. 2 λ(y)fi (y) y s. λ(y)∆fi (y) = y y λ(y)fi (y (i) ) − y λ(y)fi (y) = − y i (y) λ(y)fi (y).t. The optimization problem becomes a single variable quadratic program in δ: max s. Below we also make the substitution vi (y) = w fi (y) + problem for λ: max y to get the optimization 1 λ(y)vi (y) − C 2 λ(y) = 0. we have the simplex constraint. Then we have δ = λ(y ) = −λ(y ) in order to sum to 1.1 SMO We do not need to solve the optimization subproblem above at each pass through the data. Sequential Minimal Optimization (SMO) approach takes an ascent step that modifies the least number of variables. 6. Since y λ(y) = 0. y αi (y) + λ(y) ≥ 0.t.

6.1. SOLVING THE M3 N QP

79

0.5

0

−0.5

−1

−1.5

−2 −0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

0.4

0.5

Figure 6.2: Representative examples of the SMO subproblem. Horizonal axis represents δ with two vertical lines depicting the upper and lower bounds c and d. Vertical axis represents the objective. Optimum either occurs at the maximum of the parabola if it is feasible or the upper or lower bound otherwise.

With a = vi (y ) − vi (y ), b = C||fi (y ) − fi (y )||2 , c = −αi (y ), d = αi (y ), we have: b max [aδ − δ 2 ] s.t. c ≤ δ ≤ d, 2 (6.4)

where the optimum is achieved at the maximum of the parabola a/b if c ≤ a/b ≤ d or at the boundary c or d (see Fig. 6.1.1). Hence the solution is given by simply clipping a/b: δ ∗ = max(c, min(d, a/b)). The key advantage of SMO is the simplicity of this update. Computing the coefficients involves dot products (or kernel evaluations) to compute w fi (y ) and w fi (y ) as well as (fi (y ) − fi (y )) (fi (y ) − fi (y )).

6.1.2 Selecting SMO pairs
How do we actually select such a pair to guarantee that we make progress in optimizing the objective? Note that at least one of the assignments y must violate (KKT1) or (KKT2),

80

CHAPTER 6. M3 N ALGORITHMS AND EXPERIMENTS

1. Set violation = 0. 2. For each y, 3. 4. 5. 6. 8. 9. 10. 11. 12. KKT1: If αi (y) = 0 and vi (y) > vi (y) + , Set y = y and violation = 1 and goto 7. KKT2: If αi (y) > 0 and vi (y) < vi (y) − , Set y = y and violation = 2 and goto 7. For each y = y , If violation = 1 and αi (y) > 0, Set y = y and goto 13. If violation = 2 and vi (y) > vi (y ), Set y = y and goto 13.

7. If violation > 0,

13. Return y and y . Figure 6.3: SMO pair selection. because otherwise αi is optimal with respect to the current α−i . The selection algorithm is outlined in Fig. 6.3. The first variable in the pair, y , corresponds to a violated condition, while the second variable, y , is chosen to guarantee that solving Eq. (6.3) will result in improving the objective. There are two cases, corresponding to violation of KKT1 and violation of KKT2. Case KKT1. αi (y ) = 0 but vi (y ) > vi (y ) + . This is the case where i, y is a not support vector but should be. We would like to increase αi (y ), so we need αi (y ) > 0 to borrow from. There will always be a such a y since
y

αi (y) = 1 and αi (y ) = 0.

Since vi (y ) > vi (y ) + , vi (y ) > vi (y ) + , so the linear coefficient in Eq. (6.4) is a = vi (y ) − vi (y ) > . Hence the unconstrained maximum is positive a/b > 0. Since the upper-bound d = αi (y ) > 0, we have enough freedom to improve the objective. Case KKT2. αi (y ) > 0 but vi (y ) < vi (y ) − . This is the case where i, y is a support vector but should not be. We would like to decrease αi (y ), so we need vi (y ) > vi (y ) so that a/b < 0. There will always be a such a y since vi (y ) < vi (y ) − . Since the

6.1. SOLVING THE M3 N QP

81

lower-bound c = −αi (y ) < 0, again we have enough freedom to improve the objective. Since at each iteration we are guaranteed to improve the objective if the KKT conditions are violated and the objective is bounded, we can use the SMO in the block-coordinate ascent algorithm to converge in a finite number of steps. To the best of our knowledge, there are no upper bounds on the speed of convergence of SMO, but experimental evidence has shown it a very effective algorithm for SVMs [Platt, 1999]. Of course, we can improve the speed of convergence by adding heuristics in the selection of the pair, as long as we guarantee that improvement is possible when KKT conditions are violated.

6.1.3 Structured SMO
Clearly, we cannot perform the above SMO updates in the space of α directly for the structured problems, since the number of α variables is exponential. The constraints on µ variables are much more complicated, since each µ participates not only in non-negativity and normalization constraints, but also clique-agreement constraints. We cannot limit our ascent steps to changing only two µ variables at a time, because in order to make a change in one clique and stay feasible, we need to modify variables in overlapping cliques. Fortunately, we can perform SMO updates on α variables implicitly in terms of the marginal dual variables µ. The diagram in Fig. 6.1.3 shows the abstract outline of the algorithm. The key steps in the SMO algorithm are checking for violations of the KKT conditions, selecting the pair y and y , computing the corresponding coefficients a, b, c, d and updating the dual. We will show how to do these operations by doing all the hard work in terms of the polynomially many marginal µi variables and auxiliary “max-marginals” variables.

Structured KKT conditions As before, we define vi (y) = w fi (y) + i (y). The KKT conditions are, for all y: αi (y) = 0 ⇒ vi (y) ≤ vi (y); αi (y) > 0 ⇒ vi (y) ≥ vi (y). (6.5)

82

CHAPTER 6. M3 N ALGORITHMS AND EXPERIMENTS

select & lift

SMO update

project

Figure 6.4: Structured SMO diagram. We use marginals µ to select an appropriate pair of instantiations y and y and reconstruct their α values. We then perform the simple SMO update and project the result back onto the marginals.

Of course, we cannot check these explicitly. Instead, we define max-marginals for each clique in the junction tree c ∈ T (i) and its values yc , as: vi,c (yc ) = max [w fi (y) + i (y)],
y∼yc

αi,c (yc ) = max αi (y).
y∼yc

We also define vi,c (yc ) = maxyc =yc vi,c (yc ) = maxy∼yc [w fi (y) + i (y)]. Note that we do not explicitly represent αi (y), but we can reconstruct the maximum-entropy one from the marginals µi by using Eq. (5.13). Both vi,c (yc ) and αi,c (yc ) can be computed by using the Viterbi algorithm (one pass propagation towards the root and one outwards from the root [Cowell et al., 1999]). We can now express the KKT conditions in terms of the maxmarginals for each clique c ∈ T (i) and its values yc : αi,c (yc ) = 0 ⇒ vi,c (yc ) ≤ vi,c (yc ); αi,c (yc ) > 0 ⇒ vi,c (yc ) ≥ vi,c (yc ). (6.6)

Theorem 6.1.1 The KKT conditions in Eq. (6.5) and Eq. (6.6) are equivalent. Proof: Eq. (6.5) ⇒ Eq. (6.6). Assume Eq. (6.5). Suppose, we have a violation of KKT1: for some c, yc , αi,c (yc ) = 0, but vi,c (yc ) > vi,c (yc ). Since αi,c (yc ) = maxy∼yc αi (y) = 0,

vi (y) ≤ vi (y).c (yc ) implies the opposite: there exists y ∼ yc such that vi (y) > vi. y(a) ∼ ya with vi (y(a) ) = vi.c (yc ) > 0. a contradiction. (6.B (yB ).b (yb ).c (yc ). B} be a partition of the nodes T (i) (cliques of C (i) ) resulting from removing the edge between a and b such that a ∈ A and b ∈ B.6). Now suppose we have a violation of KKT2: for some i.c (yc ) ≥ vi. then αi (y) > 0. Then vi (y) < vi (y).6. which also implies vi (y) > vi (y). a contradiction.c (yc ) = 0.s (ys ) . But αi (y) = c∈T (i) c∈S (i) µi.c (yc ). with a separator s. vi. Assume Eq. Let {A. By Eq.c (yc ) µi. ∀y ∼ yc . αi (y) > 0. Hence. Since vi. (6.c (yc ) > 0.c (yc ). But vi.A (yA ) and vi. and it follows that all the y-consistent marginals µi are positive as well µi. but vi (y) > vi (y). ∀c ∈ T (i) (since sum upper-bounds max). which by definition agree with y .6) ⇒ Eq. hence vi. Take any maximizer. ∀c ∈ T (i) . cliques a and b. vi (y) ≥ vi (y).a∪b (ya∪b ) = vi. That is.b (yb ) ≥ vi. then maxy vi (y ) = vi. ∀c ∈ T (i) .6). αi (y) = 0. We denote the two subsets of an assignment y as yA and yB (with overlap at ys ).c (yc ) > 0. and show that vi.b (yb ). For that y. (6.c (yc ). we consider any two adjacent nodes in the tree T (i) . Now suppose we have a violation of KKT2: for some y.c (yc ) ≥ vi. Note that trivially maxy vi (y ) = max(vi. where vi.c (yc ). ∀c ∈ T (i) . αi. We will show that vi (y) = vi.1. By chaining this equality from the root of the tree to all the leaves. vi.c (yc )) for any clique c and clique assignment yc . Hence all the y-consistent αi maxmarginals are positive αi.c (yc ) = vi (y) > vi (y) > vi. . But αi.c (yc ) is the optimal value. Since αi (y) > 0.a (ya ) and any maximizer y(b) ∼ yb with vi (y(b) ) = vi. we get vi (y) = vi. ∀c ∈ T (i) . ∀c ∈ T (i) .c (yc ). but vi.c (yc ) > vi.c (yc ) for any c.c (yc ). (6.a (ya ) ≥ vi.5).c (yc ). a contradiction.a (ya ) = vi.6). vi. (6. SOLVING THE M3 N QP 83 then αi (y) = 0. Suppose we have a violation of KKT1: for some y. by Eq. (6. we know that all the y-consistent αi max-marginals are positive αi. ∀c ∈ T (i) . yc . To show that this. This means that y is the optimum of vi (·). We need to introduce some more notation to deal with the two parts of the tree induced by cutting the edge between a and b. yc ∼ y.5). by Eq.A (yA ) + vi. Eq.B (yB ) only count the contributions of their constituent cliques.c (yc ) < vi. if vi. then we cannot have αi. ∀y ∼ yc . (6.c (yc ) > 0 implies there exists y ∼ yc such that αi (y) > 0. But by Eq. The value of an assignment vi (y) can be decomposed into two parts: vi (y) = vi. a contradiction. but vi (y) < vi (y).c (yc ) > 0.c (yc ). ∀y ∼ yc .5).c (yc ) > vi. so if all the y-consistent marginals are positive.

c (yc ) > 0 and vi (y ) < vi. This is the case where i. the coefficients for the onevariable QP in Eq. so the linear coefficient in Eq. y . y is a not support vector but should be. Note that vi (y(s) ) = vi (yA ) + vi (yB ) = vi.s (ys ) since they are optimal as we said above. so αi (y ) = αi. M3 N ALGORITHMS AND EXPERIMENTS on the intersection s.3) will result in improving the objective. This is the case where i. b = C||fi (y )−fi (y )||2 . Hence the unconstrained maximum is positive a/b > 0. 6. so vi (y ) = vi.c (yc ) + > vi (y ) + and αi (y ) = 0.c (yc ) > vi. while the second variable.c (yc ) > 0.c (yc )+ . We decompose the two associated values into the corresponding parts: vi (y(a) ) = vi (yA )+vi (yB ) and vi (y(b) ) = vi (yA )+vi (yB ).4) is a = vi (y ) − vi (y ) > . we have enough freedom to improve the objective. y is a support vector but should not be. we enforce approximate KKT conditions in the algorithm in Fig. αi. we will select the first variable in the pair.c (yc ) > 0 but vi.c (yc )− . There will always be a such a y since . We have set y = arg maxy∼yc αi (y). Now we create another assignment that clamps the value of both a and b: y(a∪b) = yA ∪ yB . y . vi (y ) ≥ vi (y ) + .84 CHAPTER 6.c (yc ) > vi. since we essentially fixed the intersection s and maximized over the rest of the variables in A and B separately. corresponding to a violated condition. Hence we have vi (yA ) + vi (yB ) = vi (yA ) + vi (yB ) ≥ vi (yA ) + vi (yB ) which implies that vi (yA ) ≥ vi (yA ) and vi (yB ) ≥ vi (yB ).c (yc ) − < vi (y ) − . There will always be a such a y (with yc = yc ) since y αi (y) = 1 and αi (y ) = 0. αi. d = αi (y ). so we need αi (y ) > 0 to borrow from. αi (y ) > 0.c (yc ) = 0 but vi. We have two cases.s (ys ).c (yc ) < vi. to guarantee that solving Eq. so we need vi (y ) > vi (y ) so that a/b < 0.5.b (yb ) ≥ vi. corresponding to violation of KKT1 and violation of KKT2. We create a new assignment that combines the best of the two: y(s) = yA ∪ yB .a (ya ) = vi. (a) (b) (a) (b) (b) (a) (a) (b) (b) (a) (a) (a) (b) (b) (a) (b) (a) (b) (a) (a) (b) (b) Structured SMO pair selection and update As in multi-class problems. The value of this assignment is optimal vi (y(a∪b) ) = vi (yA ) + vi (yB ) = vi (y(a) ) = vi (y(b) ). Case KKT1. Having selected y and y . As before. (6. which guarantees that for yc = arg maxy∼yc αi (y). We can find one by choosing yc for which αi. c = −αi (y ). We would like to decrease αi (y ). Since the upper-bound d = αi (y ) > 0.c (yc ) < vi.4) are a = vi (y )−vi (y ). Since vi (y ) ≥ vi (y ) + . (6. (6. Case KKT2. We have set y = arg maxy∼yc vi (y). Now vi. We would like to increase αi (y ).

Set yc = arg maxy∼yc αi (y) and goto 13. y = arg maxy∼yc vi (y) and violation = 1 and goto 7. y = arg maxy∼yc αi (y) and violation = 2 and goto 7. 7. Set yc = yc . each character was rasterized into an image of 16 by 8 binary .c (yc ) = µi. 5. 4.c (yc ). and the change is very simple: µi. For each yc = yc . 10. yc 3. If violation = 2 and vi. from 150 human subjects. and vi. Set violation = 0.c (yc ) > vi.c (yc ) > 0.c (yc ) + . 2.c (yc ) > 0. we need to project this change onto the marginal dual variables µi .c (yc ) + δ1 c ∼ y ) − δ1 c ∼ y ). Set yc = yc . 11. from the data set collected by Kassel [1995].c (yc ) = 0. KKT1: If αi. I(y I(y 6. vi (y ) > vi (y ) − . Each word was divided into characters. We can find one by choosing yc for which vi. If violation = 1 and αi. which guarantees that for yc = arg maxy∼yc vi (y).c (yc ) > vi.2. 12. Having computed new values αi (y ) = αi (y ) + δ and αi (y ) = αi (y ) − δ. 6. Return y and y .2 Experiments We selected a subset of ∼ 6100 handwritten words.c (yc ) − . with average length of ∼ 8 characters.c (yc ) < vi. 13.5: Structured SMO pair selection. KKT2: If αi. and vi. If violation > 0.c (yc ) − .c (yc ) > vi. 9.6. Since the lowerbound c = −αi (y ) < 0. Figure 6. EXPERIMENTS 85 1. Set y = arg maxy∼yc vi (y) and goto 13. The only marginal affected are the ones consistent with y and/or y . For each c ∈ T (i) . again we have enough freedom to improve the objective. vi (y ) < vi (y ) − . 8.

For the sequence approaches (CRFs and M3 ). Logistic regression and CRFs are both trained by maximizing the conditional likelihood of the labels given the features. Our features for each label Yj are the corresponding image of ith character. a label of an individual character to Yj .6(b).6(a). (2. multi-class SVMs as described in Eq. M3 N ALGORITHMS AND EXPERIMENTS 0.2 0. are averages over the 10 folds. using a zero-mean diagonal Gaussian prior over the parameters. The other methods are trained by margin maximization. which do not consider the correlation between neighboring characters — logistic regression. .15 0. and a labeling for a complete word to Y. pixels. (b) OCR: Average percharacter test error for logistic regression. . . Each label Yj takes values from one of 26 classes {a. 6.35 linear Test error (average per-character) 0. the image for each word corresponds to x. multiclass SVMs.25 0. .1 and 1.9). quadratic. and sequence approaches — CRFs. CRFs. (See Fig. The accuracy results. . and our proposed M3 networks. 6. and M3 Ns.3 0. We implemented a selection of state-of-the-art classification algorithms: independent label approaches. with a standard deviation between 0. we used an indicator basis function to represent the correlation between Yj and Yi+1 . and one-against-all SVMs (whose performance was slightly lower than multi-class SVMs).05 0 Log-Reg CRF quadratic cubic m SVM M^3N (a) (b) Figure 6.) In our framework.6: (a) 3 example words from the OCR data set. and cubic kernels.86 CHAPTER 6. using linear. The data set is divided into 10 folds of ∼ 600 training and ∼ 5500 testing examples.1 0. z}. summarized in Fig.

g.. it has yet to be shown to learn faster than Structured SMO. empirically. are about comparable to those of multiclass SVMs. 1998] for large-margin classifiers have a similar online optimization scheme. our sequence model significantly outperforms other approaches by incorporating . In our experiments with the OCR task. 2004]. 1998] and voted-perceptron algorithms [Freund & Schapire. Second. 6. 1997] algorithm has been adopted to solve our structured QP for max-margin estimation [Bartlett et al. Collins [2001] have applied voted-perceptron to structured problems in natural language. 6.4 Conclusion In this chapter. 6.6(b) shows two types of gains in accuracy: First. Using these high-dimensional feature spaces in CRFs is not feasible because of the enormous number of parameters.3. For comparison.6. margin-based methods achieve a very significant gain over the respective likelihood maximizing methods. the Exponentiated Gradient [Kivinen & Warmuth. RELATED WORK 87 For margin-based methods (SVMs and M3 ). less passes (about 30-40) are needed for our algorithm than in the perceptron literature. Once we use cubic kernels our error rate is 45% lower than CRFs and about 33% lower than the best previous approach. we were able to use kernels (both quadratic and cubic were evaluated) to increase the dimensionality of the feature space. We used the structured SMO algorithm with about 30-40 iterations through the data. the error rate of our method using linear features is 16% lower than that of CRFs.. we address the large (though polynomial) size of our quadratic program using an effective optimization procedure inspired by SMO. a larger training set). it seems that. particularly in the early iterations through the dataset. by using sequences.. Although head-to-head comparisons have not been performed. Interestingly. and about the same as multi-class SVMs with cubic kernels. Recently. although using a different setup (e. we obtain another significant gain in accuracy. the previously published results.3 Related work The kernel-adatron [Friess et al. by using kernels. Fig. Although the EG algorithm has attractive convergence properties.

Overall.88 CHAPTER 6. we apply this framework to important classes of Markov networks for spatial and relational data. we believe that M3 networks will significantly further the applicability of high accuracy margin-based methods to real-world structured data. M3 N ALGORITHMS AND EXPERIMENTS high-dimensional decision boundaries of polynomial kernels over character images while capturing correlations between consecutive characters. . In the next two chapters.

webpage 89 . 1994]. In some cases. in which connected (“associated”) variables tend to have the same label.Chapter 7 Associative Markov networks In the previous chapter. which allow very efficient inference and learning. which is NP-hard in general networks suitable in a broad range of practical Markov network structures. leading to computationally intractable networks. which is usually highly interconnected. For example. such as the quad-tree model used for image segmentation [Bouman & Shapiro. In many cases. we considered applications of sequence-structured Markov networks. contains networks of discrete variables with K labels and arbitrary-size clique potentials with K parameters that favor the same labels for all variables in the clique. Such positive interactions capture the “guilt by association” pattern of reasoning present in many domains.g.. In other cases (e. In this chapter. The chief computational bottleneck in applying Markov networks for other large-scale prediction problems is inference. subject to the constraint that inference on these networks is tractable. called associative Markov networks (AMNs). the topology of the Markov network does not allow tractable inference. however. the network structure can mirror the hyperlink graph. AMNs are a natural fit object recognition and segmentation. in hypertext. including grid-topology networks [Besag. the network structure is learned. 1986]. This subclass. 2001]). we show that optimal learning is feasible for an important subclass of Markov networks — networks with attractive potentials. a tractable structure is determined in advance. [Bach & Jordan. One can address the tractability issue by limiting the structure of the underlying network.

For binary-valued Potts models. By contrast.. a common approach is to use a generalized Potts model [Potts. We use a linear program relaxation of this integer program in the min-max formulation. this linear program provides exact answers.1 Associative networks Associative interactions arise naturally in the context of image processing. while balancing the spatial coherence modeled by the AMN. other learning tasks (e. ASSOCIATIVE MARKOV NETWORKS classification. We show that. In this setting. For the non-binary case (K > 2). The MAP problem can naturally be expressed as an integer programming problem. the MAP problem . our models can be learned efficiently and at run-time. In the max-margin estimation framework. where nearby pixels are likely to have the same label [Besag. which penalizes assignments that do not have the same label across the edge: φij (k. Thus. Boykov et al. the MAP problem can be solved exactly for this class of models in polynomial time. We present an AMN-based method for object segmentation of complex from 3D range data. The proposed learning formulation effectively and directly learns to exploit a large set of complex surface and volumetric features. ∀k = l and φij (k. the approximate linear program is not guaranteed to be optimal but we can bound its relative error. for associative Markov networks of over binary variables (K = 2). l) = λij . k) = 1.g. and many other applications. For L > 2. Greig et al. 1952]. where λij ≤ 1. To our knowledge. rather than just the MAP. Our empirical results suggest that the solutions of the resulting approximate maxmargin formulation work well in practice.. By constraining the class of Markov networks to AMNs. maximizing the conditional likelihood of the target labels given the features) require that we compute the posterior probabilities of different label assignments. our method is the first to allow training Markov networks of arbitrary connectivity and topology. [1989] show that the MAP problem can be formulated as a min-cut in an appropriately constructed graph. 7. 1999b]. scale up to tens of millions of nodes and edges.90 CHAPTER 7. the inference subtask is one of finding the best joint (MAP) assignment to all of the variables in a Markov network. 1986.

2003]. Definition 7. This additional flexibility is important in many domains.. . 1999b. but a procedure based on a relaxed linear program guarantees a factor 2 approximation of the optimal solution [Boykov et al. k) = λc (k) ≥ 1 and the rest of the entries are set to 1. consider the problem of predicting whether two proteins interact [Vazquez et al. . AMNs admit non-pairwise interactions between variables. and µc (k) for each clique c . In this case. l) = 1. Our associative potentials extend the Potts model in several ways. More formally. k). where λij (k) ≥ 1. . I(y gc (k) ≥ 0. This type of transitivity could be modeled by a ternary clique that has high λ for the assignment with all interactions present. ∀k = l. Importantly. .. µim ). In particular. AMNs allow us to encode the pattern of (soft) transitivity present in many domains. Kleinberg & Tardos. In a second important extension. . For example. . . 1999]. where we have a “marginal” variable µv (k) for each node v ∈ V and each label k. LP INFERENCE 91 is NP-hard. .2. where V are the nodes and C are the cliques of the graph G. the clique potentials are constrained to have the same type of structure as the edge potentials: There are K parameters φc (k.1. using this additional expressive power. k) = λij (k). A set of potentials φ(y) is associative if φ(y) = eg(y) and g(y) is associative. as different labels can have very diverse affinities. .7. For example. foreground pixels tend to have locally coherent values while background is much more varied. AMNs allow different labels to have different attraction strength: φij (k. and φij (k. ∀c ∈ C\V. we define associative functions and potentials as follows.2 LP Inference We can write an integer linear program for the problem of finding the maximum of an associative function g(y).1 A function g : Y → IR is associative for a graph G over K-ary variables if it can be written as: K K g(y) = v∈V k=1 gv (k)1 v = k) + I(y c∈C\V k=1 gc (k)1 c = k. . with potentials over cliques involving m variables φ(µi1 . which indicates whether node v has value k. . 7. . this probability may increase if they both interact with another protein.

∀c ∈ C \ V. for any associative function g.1 for the proof. In the appendix. .2. In fact.g. Theorem 7.1) s. 1}. This works because the coefficient gc (k) is non-negative and we are maximizing the objective function. µc (k) ∈ {0. This result states that the MAP problem in binary AMNs is tractable. (7. the linear relaxation of Eq. v∈c Note that we substitute the constraint µc (k) = µv (k) by linear inequality con- straints µc (k) ≤ µv (k). which is equivalent to µc (k) = v∈c µv (k).2 If K > 2. µc (k) ≤ µv (k). the linear relaxation of Eq. for any associative function g. (7.2.t. v ∈ c.1 If K = 2.92 CHAPTER 7.2. which represents the event that all nodes in the clique c have label k: K K max v∈V k=1 µv (k)gv (k) + c∈C\V k=1 µc (k)gc (k) K (7. when µv (k) are binary.1).1) has an integral optimal solution. we never encountered fractional solutions. the linear relaxation of Eq. µc (k) = minv µv (k) . (where the constraints µc (k) ∈ {0. regardless of network topology or clique size. k=1 µv (k) = 1. ∀c ∈ C.1) has a solution that is larger than the solution of the integer program by at most the number of variables in the largest clique. we also show that the approximation ratio of the rounding procedure is the inverse of the size of the largest clique (e. ASSOCIATIVE MARKOV NETWORKS (containing more than one variable) and label k. Hence at the optimum. it seems that in real data such symmetries are often broken. Although artificial examples with fractional solutions can be easily constructed by using symmetry. k. k. these LPs can produce fractional solutions and we use a rounding procedure to get an integral solution. (7. ∀v ∈ V. See Appendix A. is guaranteed to produce an integer solution when a unique solution exists. It can be shown that in the binary case. 1 2 for pairwise networks). 1} are replaced by µc (k) ≥ 0).. In the non-binary case (L > 2). Theorem 7. in all our experiments with L > 2 on real data.

1) is: max v∈V [µv (1)gv (1) + µv (2)gv (2)] + uv∈E [µuv (1)guv (1) + µuv (2)guv (2)]. and then how to combine the two. The reason for that is that the final min-cut graph must consist of only positive weights. (7.3 Min-cut inference We can also use efficient min-cut algorithms to perform exact inference on the learned models for K = 2 and approximate inference for K > 2. An .7. (7. We then look at ∆v = gv (1) − gv (2). We will show how to deal with the node terms (those depending only on a single variable) and the edge terms (those depending on a pair of variables). Such a term corresponds to the node potential contribution to our objective function for node v. min − v∈V [µv (1)gv (1) + µv (2)gv (2)] − uv∈E [µuv (1)guv (1) + µuv (2)guv (2)]. First.1 Graph construction We construct a graph in which the min-cut will correspond to the optimal MAP labeling for the above objective. and create an edge of weight |∆v | from v to either 1 or 2.3. For pairwise.2) 7. MIN-CUT INFERENCE 93 7. we focus on the pairwise AMN case. and the V2 set will correspond to label 2.3. we recast the objective as minimization by simply reversing the signs on the value of each θ. V2 ) cut. For each node term corresponding to node v we add a vertex v to the min-cut graph. the objective of the integer program in Eq.3) The graph will consist of a vertex for each node in the AMN. binary AMNs. [1999a] to perform (approximate) inference in the general multi-class case. For simplicity. depending on the sign of ∆v . the V1 set will correspond to label 1. along with the 1 and 2 terminals. (7. In the final (V1 . Node terms Consider a node term −µv (1)gv (1) − µv (2)gv (2). We first consider the case of binary AMNs. and later show how to use the local search algorithm developed by Boykov et al.

since for each introduced vertex we have only one edge of positive weight to either 1 or 2.3. ASSOCIATIVE MARKOV NETWORKS 1 v 2 1 u v 1 v 2 2 Figure 7. example is presented in Fig. such cuts will be possible.94 CHAPTER 7. When looking at edge terms in isolation. but when we combine the graphs for node terms and edge terms. We can take the individual graphs we created for node and edge terms and merge them . Edge Terms Now consider an edge term of the form −µuv (1)guv (1) − µuv (2)guv (2).3.3. 7. and we would always choose not to cut any edges. Observe what happens when both nodes are on the V2 side of the cut: the value of the min-cut is guv (1). From Fig.1. which must be less than guv (2) or the min-cut would have placed them both on the 1 side. We will connect vertex u to 1 with an edge of weight guv (1). Fig.1 shows an example. the graph construction above would add an edge from each node to its more likely label. we would simply get a cut with cost 0. To construct a min-cut graph for the edge term we will introduce two vertices u and v. Thus if we run min-cut.1. connect v to 2 with an edge of weight guv (2) and connect u to v with an edge of weight guv (1) + guv (2). a cut that places each node in different sets will not occur.1: Min-cut graph construction of node (left) and edge (right) terms. we see that if the AMN consisted of only node potentials. 7. 7.

[1999a]. One of the most effective algorithms for minimizing energy functions like ours is the α-expansion algorithm proposed by Boykov et al. .3. and finds the new labeling whose objective has the lowest value. The algorithm performs a series of “expansion” moves each of which involves optimization over two labels. Another labeling µ is called an “α-expansion” move (following Boykov et al. 7. MIN-CUT INFERENCE 95 by adding edge weights together (and treating missing edges as edges with weight 0). or switch to the label α.) In other words. and it can be shown that it converges to within a factor of 2 of the global minimum. It terminates when there is no α-expansion move for any label k that has a lower objective than the current labeling (Fig. . The α-expansion algorithm cycles through all labels k in either a fixed or random order. or change to k. a k-expansion from a current labeling allows each label to either stay the same. K.2 Multi-class case The graph construction above finds the best MAP labeling for the binary case.3. [1999a]) from µ if µv = k implies µv = µv (where µv is the label of the node v in the AMN.2). Since our MAP-inference objective is simply a sum of node and edge terms. In this new binary problem we will let label 1 represent a node keeping its current label and label 2 will denote a node taking on the new label k. but in practice we would often like to handle multiple classes in AMNs. The key part of the algorithm is computing the best α-expansion labeling for a fixed k and a fixed current labeling µ. It can be shown that the resulting graph will represent the same objective (in the sense that running min-cut on it will optimize the same objective) as the sum of the objectives of each graph. Expansion Algorithm Consider a current labeling µ and a particular label k ∈ 1. In order to construct the right coefficients . . The min-cut construction from earlier allows us to do exactly that since an α-expansion move essentially minimizes a MAP-objective over two labels: it either allows a node to retain its current label. 7.7. . merging the node and edge term graphs will result in a graph in which min-cut will correspond to the MAP labeling.

As noted in Boykov et al. ASSOCIATIVE MARKOV NETWORKS 1. we set θij = θiji j . [1999a] and as we have observed in our experiments. and θi1 = θi . .2 If E(ˆ) < E(µ). and θyi denotes the coefficient in the multiclass AMN MAP objective. we can apply the min-cut construction from before to get the optimal labeling within one α-expansion from the current one. ˆ 3. j) ∈ E in which exactly one of the α. let θik and θij denote the node and edge coefficients associated with the new binary objective: ◦ Node Potentials For each node i in the current labeling whose current label is not α. Set success := 0 3. we add θij . j) ∈ E whose nodes have labels different from α. Return µ Figure 7. K} 3.k for the new binary objective we need to consider several factors. Below.2: α-expansion algorithm k. y α we let θi0 = θi i . with weights θij = θij . α. where yi denotes the current label of node i. the algorithm terminates only after a few iterations with most of the improvement occurring in the first 2-3 expansion moves.α nodes has label α in the current labeling. For each edge (i.1 Compute µ = arg min −g(µ ) among µ within one α-expansion of µ. . After we have constructed the new binary MAP objective as above. If success = 1 goto 2. 5.y node whose label is different from α.α 1 we add a new edge potential. and if the two nodes currently have 0 different labels we let θij = 0. . to the node potential θi1 of the y . set µ := µ and success := 1 µ ˆ 4. For each label k ∈ {1. Note that we ignore nodes with label α altogether since an α-expansion move cannot change their label. . Begin with arbitrary labeling µ 2. ◦ Edge Potentials For each edge (i. Veksler [1999] shows that the α-expansion algorithm converges in O(N ) iterations where N is the number of nodes. If the two nodes of the edge 0 currently have the same label.96 CHAPTER 7.

MAX-MARGIN ESTIMATION 97 7.3 w f (x. ˙ Definition 7.4. In practice.2 Let f be the subset of basis functions f with support only on singleton cliques: ˙ f = {f ∈ f : ∀ x ∈ X . and any combination of basis functions with support only on singleton cliques is also associative. We must make similar associative assumption on the loss function in order to guarantee that the LP inference can handle it. all basis functions evaluate to 0 for assignments where the values of the nodes are not equal and are non-negative for the assignments where the values of the nodes are equal. Recall that this implies that for cliques larger than one. w} = w be the corresponding f subsets of parameters. y) whenever Assumption 7.1 ¨ holds and w ≥ 0. y) is associative for G (i) for all i. fc (xc . it is useful to separate the basis functions with support only on nodes from those with support on larger cliques.4. It is easy to verify that any non-negative combination of associative functions is associative. which we use in generalization bounds and experiments.1 Basis functions f are component-wise associative for G(x) for any (x.4. y) is associative. . y) is associative: Assumption 7.4. c ∈ C(G(x)). To ensure that w f (x. ˙ ˙ ¨ Let ¨ = f \ f be the rest of the basis functions. is associative. yc ) = 0}. y ∈ Y. y) is associative for G(x) for any (x. y(i) . Assumption 7.4 The loss function (x(i) . so we have: Lemma 7. y). Let {w.7.4. and the Hamming loss. |c| > 1.4.4 Max-margin estimation The potentials of the AMN are once again log-linear combinations of basis functions. this restriction is fairly mild. We will need the following assumption to ensure that w f (x.

c∈C (i) .3 for derivation).c (k) ≤ µi. ∀c ∈ C (i) k=1 (i) µi. However. and not surprisingly.v (k).2. k.4). k (i) C ˙ µi.c (k) f i. there are marginals µ for each node and clique. k. w ∆fi.v (k) − c⊃v mi. the constraints are different.4 and some algebra (see Appendix A. v ∈ c. In the dual. the basic structure of the first two sets of constraints remains the same: we have local margin requirements and “credit” passed around through messages mi.v i. for each value k.c (k). c ∈ C (i) \ V (i) .t.c (yc )/|c|.v (k)∆fi. ¨ ∀i. ∀i.v∈V (i) .c.v . (5. k) and i. \ V . .t. The extra constraints are due to the associativity constraints on the resulting model.c (k) i.c. ∀i. similar to Eq.c (k) = fi.c (k.v (k) ≥ v∈c − ξi. (i) ¨ f mi. we have the following max-margin QP for AMNs: min s.v∈V (i) (7. k. .v (k) = 1.v (k). .v (k) ≥ mi. ∀v ∈ V (i) . (7.c (k)∆¨i. v ∈ V (i) .4. A.c (k.c (k) ≥ 0. c ∈ C (i) \ V (i) . ∀i. While this primal is more complex than the regular M3 N factored primal in Eq. .v (k) ≥ −w ¨i. k. v ∈ c.1 and 7. k s.c. where fi. µi.v (k) i.98 CHAPTER 7. ∀c ∈ C . .c.4) i. . ASSOCIATIVE MARKOV NETWORKS Using the above Assumptions 7.4) (see derivation in Sec.c∈C (i) . (7. The dual of Eq. µi. . ¨ w ≥ 0. ν ≥ 0.c (k) − 2 i.c (k) = i. k).1).c (k) + f ∀i.3) is given by: 2 2 max i.2. (5. k C µi. k.v (k) − 2 K ν+ ¨ µi.4.12). are essentially the constraints from the inference LP relaxation in Eq. ¨ w ∆¨i. ∀i. . 1 ||w||2 + C 2 ξi. .

these weights tend to do well in practice.v (k). For K > 2. Thus. Each reading was a point in 3D space.c∈C (i) . the only available information is the location of points. building.c (k) f ¨ The ν variables simply ensure that w are positive (if any component ¨ i.c (k)∆¨i.c (k).v∈V (i) . The data consists of around 35 million points.¯(k). Note that the objective can be written in terms of dot products of node basis func˙ ˙ v ¯ tions ∆fi. Terrain classification is very useful for autonomous mobile robots in real-world environments for path planning. represented as 3D coordinates in an absolute frame of reference Fig. represented by its (x. 7. ¨ w=ν+ ¨ i. ξ for the original QP may no longer be feasible. Thus. the optimal choice w.v (k)∆fi.7. k ˙ µi. k is negative. the LP inference is exact.c∈C (i) .4) learns exact max-margin weights for Markov networks of arbitrary topology. However. The Stanford Segbot Project1 has provided us with a laser range maps of the Stanford campus collected by a moving robot equipped with SICK2 laser sensors Fig. (7. so they can be kernelized. so that Eq.c (k)∆¨i.v (k) ∆fj. and as a pre-processing step for other perceptual tasks. Our task is to classify the laser range points into four classes: ground. Thus. maximizing the objective will force the corresponding component of ν to cancel ¨ it out). y. leading to a different choice of weights. tree. the only available information is the location of points.5. z) coordinates in an absolute frame of reference. 7. Unfortunately. k µi. For K = 2.5. target detection. f µi. EXPERIMENTS 99 The dual and primal solutions are related by ˙ w= i.5 Experiments We applied associative Markov networks to the task of terrain classification. the linear relaxation leads to a strengthening of the constraints on w by potentially adding constraints corresponding to fractional assignments as in the case of untriangualated networks. . 7. as our experiments show.5. which was fairly noisy because of localization errors. 1 Many thanks to Michael Montemerlo and Sebastian Thrun for sharing the data. the edge basis functions cannot be kernelized because of the non-negativity constraint.

After we do that.3: Segbot: roving robot equipped with SICK2 laser sensors. The features we use require pre-processing to infer properties of the local neighborhood of a point. Since classifying ground points is trivial given their absolute z-coordinate (height). We use a number of features derived from the cube such as the percentage of points in the central column.100 CHAPTER 7. and compute the percentage of points lying in the various sub-cubes. we are left with approximately 20 million non-ground points. the central plane. For each point we sample 100 points in a cube of radius 0. such as how planar the neighborhood is. or how much of the neighbors are close to the ground. We run PCA on these points to get the plane of maximum variance (spanned by the first two principal components). Each point is represented simply as a location in an absolute 3D coordinate system. the outside corners. ASSOCIATIVE MARKOV NETWORKS Figure 7. Our second type of feature is based on a column . We then partition the cube into 3 × 3 × 3 bins around the point. we classify them deterministically by thresholding the z coordinate at a value close to 0.5 meters. since scans tend to be sparser in regions farther from the robot. The features we use are invariant to rotation in the x-y plane. and shrubbery. oriented with respect to the principal plane. as well as the density of the range scan. These features capture the local distribution well and are especially useful in finding planes. etc. Our first type of feature is based on the principal plane around it.

augmented with a quadratic kernel. This feature is especially useful in classifying shrubbery. 7. All methods use the same set of features. The first model is a multi-class SVM with a quadratic kernel over the above features.7. We considered three different models: SVM.4: 3D laser scan range map of the Stanford Quad. around each point. EXPERIMENTS 101 Figure 7. we also use an indicator feature of whether or not a point lies within 2m of the ground. which extends vertically to include all the points in a “column”.5.25 meters. between 2m and 2. We then compute what percentage of the points lie in various segments of this vertical column (e. a tree. For training we select roughly 30 thousand points that represent the classes well: a segment of a wall. 7.. Voted-SVM and AMNs. right panel and Fig. top panel) achieves reasonable performance .g. We take a cylinder of radius 0. some bushes. This model (Fig.5m). Finally.7.5.

For example arches on buildings and other less planar regions are consistently confused for trees.7. middle panel and Fig. with associative potentials to ensure smoothness. 7. trees/green. 7. Each point is connected to 6 of its neighbors: 3 of them are sampled randomly from the local neighborhood in a sphere of radius 0.5: Terrain classification results showing Stanford Memorial Church obtained with SVM.5. middle panel) performs slightly better than SVM: for example. (Color legend: buildings/red. It is important to ensure vertical consistency since the SVM classifier is wrong in areas that are higher off the . and the other 3 are sampled at random from the vertical cylinder column of radius 0. The Voted-SVM model (Fig. ASSOCIATIVE MARKOV NETWORKS Figure 7. We improved upon the SVM by smoothing its predictions using voting. ground/gray).5m. Yet it still fails in areas like arches of buildings where the SVM classifier has a locally consistent wrong prediction. shrubs/blue.102 CHAPTER 7. it smooths out trees and some parts of the buildings. but fails to enforce local consistency of the classification predictions. Voted-SVM and AMN models. even though they are surrounded entirely by buildings. The final model is a pairwise AMN over laser scan points. For each point we took its local neighborhood (we varied the radius to get the best possible results) and assigned the point the label of the majority of its 100 neighbors.25m. in many places.

and as we can see.6: The running time (in seconds) of the min-cut-based inference algorithm for different problem sizes. which uses bidirectional search trees for augmenting paths (see Boykov and Kolmogorov [2004]). We used a publicly available implementation of the min-cut algorithm.5. bottom panel) are much smoother: for example building arches and tree trunks are predicted .5 x 10 7 Figure 7.7. We can see that the predictions of the AMN (Fig. 7. The problem size is the sum of the number of nodes and the number of edges.5 1 1. Note the near linear performance of the algorithm and its efficiency even for large models. 7. While we experimented with a variety of edge features including various distances between points.5 Problem size (nodes and edges) 2 2.5. 7. it is roughly linear in the size of the problem (number of nodes and number of edges).7. We trained the AMN model using CPLEX to solve the quadratic program. The implementation is largely dominated by I/O time. The performance is summarized in Fig. we found that even using only a constant feature performs well. tree trunks and tree crowns look different). The inference over each segment was performed using min-cut with α-expansion moves as described above. with the actual min-cut taking less than two minutes even for the largest segment. left panel and Fig. the training took about an hour on a Pentium 3 desktop. ground (due to the decrease in point density) or because objects tend to look different as we vary their z-coordinate (for example.6. EXPERIMENTS 103 Mincut inference performance 250 200 Running time (seconds) 150 100 50 0 0 0.

our approach can also handle multi-class problems in a straightforward manner. More recently. Our definition of associative potentials below also satisfies the Kolmogorov and Zabih regularity condition for K = 2. Kleinberg and Tardos [1999] extend the multi-class Potts model to have more general edge potentials. They also provide a solution based on a relaxed LP that has certain approximation guarantees. as long as the parameters satisfy a certain regularity condition. under the constraints that negative log of the edge potentials form a metric on the set of labels. we provide an algorithm for max-margin training of associative Markov networks. In fact. our learning formulation provides an exact and tractable optimization algorithm. We also hand-labeled around 180 thousand points of the test set (Fig. Kolmogorov and Zabih [2002] showed how to optimize energy functions containing binary and ternary interactions using graph cuts.stanford. including a fly-through movie of the data. at http://ai.edu/˜btaskar/3Dmap/. 7.8) and computed accuracies of the predictions shown in Fig. Unlike their work. However. Kumar and Hebert [2003] train CRFs using a pseudo-likelihood approximation to the distribution P (Y | X) since estimating the true conditional distribution is intractable. we can extend our max-margin framework to estimate their more general potentials by expressing inference as a linear program. which was classified by pre-processing). a subclass of Markov networks that allows only positive interactions between . See more results. Moreover. 7. the structure of our potentials is simpler to describe and extend for the multi-class case.104 CHAPTER 7. Our terrain classification approach is most closely related to work in vision applying conditional random fields (CRFs) to 2D images. ASSOCIATIVE MARKOV NETWORKS correctly.9 (excluding ground.6 Related work Several authors have considered extensions to the Potts model.7 Conclusion In this chapter. as well as formal guarantees for binary classification problems. unlike their work. 7. The differences are dramatic: SVM: 68%. 7. Voted-SVM: 73% and AMN: 93%.

It would be very interesting to consider other applications.7. We present large-scale experiments with terrain segmentation and classification from 3D range data involving AMNs with tens of millions of nodes and edges. We thus provide a polynomial time algorithm which approximately solves the maximum margin estimation problem for any associative Markov network. The class of associative Markov networks appears to cover a large number of interesting applications. the margin) of our learned model. In the non-binary case. Our results in the binary case rely on the fact that the LP relaxation of the MAP problem provides exact solutions. and it is an interesting challenge to apply the algorithm to even larger models and develop efficient distributed implementations. CONCLUSION 105 related variables. we are not guaranteed exact solutions.g. or predicting links in relational data. . Our approach relies on a linear programming relaxation of the MAP problem. which is the key component in the quadratic program associated with the maxmargin formulation. To our knowledge. The min-cut based inference is able to handle very large networks. it is clear that many applications call for repulsive potentials.7. It would be interesting to see whether these bounds provide us with guarantees on the quality (e. such as extracting protein complexes from protein-protein interaction data. this algorithm is the first to provide an effective learning procedure for Markov networks of such general structure. Importantly. While clearly we cannot introduce fully general potentials into AMNs without running against the NP-hardness of the general problem. our method is guaranteed to find the optimal (margin-maximizing) solution for all binary-valued AMNs. and consider another one (hypertext classification) in the next. However. it would be interesting to see whether we can extend the class of networks we can learn effectively.. regardless of the clique size or the connectivity. despite the prevalence of fully associative Markov networks. We have explored only a computer vision application in this chapter. but we can prove constant-factor approximation bounds on the MAP solution returned by the relaxed LP.

.7: Results from the SVM. Voted-SVM and AMN models. ASSOCIATIVE MARKOV NETWORKS Figure 7.106 CHAPTER 7.

CONCLUSION 107 Figure 7.7. .8: Labeled part of the test set: ground truth (top) and SVM predictions (bottom).7.

. ASSOCIATIVE MARKOV NETWORKS Figure 7.9: Predictions of the Voted-SVM (top) and AMN (bottom) models.108 CHAPTER 7.

and the link structure is an important source of information. social networks. such as a partition into sections. However. Each document also has internal structure.Chapter 8 Relational Markov networks In the previous chapters. Most naively. hypertext has a very rich structure that this approach loses entirely. where each entity type is characterized by a different set of attributes. the entities to be labeled are related with each other in very complex ways. the labels of linked pages are highly correlated. In this chapter. Consider a collection of hypertext documents that we want to classify using some set of labels. One document has hyperlinks to others. Such data consist of entities of different types. A standard approach is to classify each entity independently. classifying each webpage solely using the words that appear on the page. Entities are related to each other via different types of links. we can use a bag of words model. cross-citations in patents and scientific papers. not just sequentially or spatially. we have seen how sequential and spatial correlation between labels can be exploited for tremendous accuracy gains. hyperlinks that emanate from the same section of the document are even 109 . we present a framework that builds on Markov networks and provides a flexible language for modeling rich interaction patterns in structured data. In many other supervised learning tasks. For example. in hypertext classification. and more. We provide experimental results on a webpage classification task. medical records. showing that accuracy can be significantly improved by modeling relational dependencies. Many real-world data sets are innately relational: hyperlinked webpages. ignoring the correlations between them. typically indicating that their topics are related.

and reference attributes E.R (for example. As we show. their attributes and relations between them. Reference attributes include a special unique key attribute E. It would also have the label attribute Doc. Doc would have a set of boolean attributes Doc. .To. .Y (for example. Doc. It thereby allows us tremendous flexibility in representing complex patterns. which describes entities. indicating the topic of the page.HasWordk ). When classifying a collection of documents. and thereby can explicitly take advantage of the correlations between the labels of related entities. The Link entity type has two attributes: Link. label attributes E. a schema specifies of a set of entity types E = {E1 . In our domain. consecutive links in a document tend to refer to documents with the same label. the use of an undirected graphical model avoids the difficulties of defining a coherent generative model for graph structures in directed models. that can potentially help us achieve better classification accuracy. We propose the use of a joint probabilistic model for an entire collection of related entities. Each type E is associated with three sets of attributes: content attributes E.From and Link.110 CHAPTER 8.HasWordk indicating whether the word k occurs on the page. both of which refer to Doc entities. these are important cues.K that identifies each . We can also capture patterns that involve groups of links: for example. Therefore. Doc. where we simultaneously decide on the class labels of all of the entities together. there are two entity types: Doc and Link. In general. we restrict label and content attributes to take on categorical values. which compactly defines a Markov network over a relational data set. which takes on a set of categorical values. rather than classifying each document separately.1 Relational classification Consider hypertext as a simple example of a relational domain. Link. En }. we can represent a pattern where two linked documents are likely to have the same topic. If a webpage is represented as a bag of words. 8. A relational domain is defined by a schema. RELATIONAL MARKOV NETWORKS more likely to point to similar documents.Label. . We introduce the framework of relational Markov networks (RMNs). For simplicity. and can easily model complex patterns over related entities. we want to provide a form of collective classification.Label). For example.X (for example.To). . The graphical structure of an RMN is based on the relational structure of the domain.

r. specifies the set of entities (nodes) and their reference attributes (edges). We will use I. This network reflects the interactions between related instances by allowing us to represent correlations between their attributes. RELATIONAL CLASSIFICATION 111 entity. We would have a probabilistic dependency from each document’s label to the words on the document. Friedman et al. 2002] are a relational extension of Bayesian networks [Pearl. For example. Given a particular instantiation graph.x. I. Getoor et al.y and I.. I. A PRM specifies a probability distribution over instantiations consistent with a given instantiation graph by specifying a Bayesian-network-like template-level probabilistic model for each entity type. Other reference attributes E.8.Label of two documents if there is a link between them. 1988]. The component I. More importantly.R to denote the content. using both the words in the title .Label) associated with it. In our hypertext example. 1998.1. specifying their labels. label and reference attributes in the instantiation I. but not their words or labels. A hypertext instantiation graph specifies a set of webpages and links between them. which we call an instantiation skeleton or instantiation graph.Label to Doc. each of these attributes would have a conditional probability distribution P (Doc. with a directed edge between Doc. Taskar et al. a PRM can represent the inter-dependencies between topics of linked documents by introducing an edge from Doc.K). An instantiation I of a schema E specifies the set of entities I(E) of each entity type E ∈ E and the values of all attributes for all of the entities.HadWordk .. an instantiation of the hypertext schema is a collection of webpages.R refer to entities of a single type E = Range(E.HasWordk | Doc.Y and I. [2001] suggest the use of probabilistic relational models (PRMs) for the collective classification task. words they contain and links between them. indicating the probability that word k appears in the document given each of the possible topic labels. Given a particular instantiation graph containing some set of documents and links. 1999. I.X. a PRM might use a naive Bayes model for words. the PRM induces a large Bayesian network over that instantiation that specifies a joint probability distribution over all attributes of all of the entities. PRMs [Koller & Pfeffer. show that this approach works well for classifying scientific documents. and a dependency from each document’s label to the labels of all of the documents to which it points.r to denote the values of those attributes. the PRM specifies a Bayesian network over all of the documents in the collection.Label and each attribute Doc.R) and take values in Domain(E . Taskar et al.

However the application of this idea to other domains. The model defines the probability of the link existence as a function of the labels of the two endpoints. This quadratic blowup occurs even when the actual link graph is very sparse. [2001] suggest an approach where we do not include direct dependencies between the labels of linked webpages. This model. In many cases. directed models such as Bayesian networks and PRMs are usually trained to optimize the joint probability of the labels and other attributes. Representing more complex patterns involving correlations between multiple edges is very difficult. In particular. The link existence model assumes that the presence of different edges is a conditionally independent event. For example. and the cyclicity problem disappears. which may or may not exist in the data. . When N is large (e.112 CHAPTER 8. if two pages point to the same page. however. RELATIONAL MARKOV NETWORKS and abstract and the citation-link structure. but rather treat links themselves as random variables. Furthermore. discriminatively trained models are more robust to violations of independence assumptions and achieve higher classification accuracy than their generative counterparts. a quadratic growth is intractable. has other fundamental limitations.g. the resulting Bayesian network has a random variable for each potential link — N 2 variables for collections containing N pages. while the goal of classification is a discriminative model of labels given the other attributes. is problematic since there are many cycles in the link graph. it is more likely that they point to each other as well. Such interactions between many overlapping triples of links do not fit well into the generative framework.. the set of all webpages). leading to cycles in the induced “Bayesian network”. labels have no incoming edges from other labels. The advantage of training a model only to discriminate between labels is that it does not have to trade off between classification accuracy and modeling the joint distribution over non-label attributes. In this link existence model. Each two pages have a “potential link”. which is therefore not a coherent probabilistic model. such as webpages. Getoor et al. Even more problematic are the inherent limitations on the expressive power imposed by the constraint that the directed graph must represent a coherent generative model over graph structures.

RELATIONAL MARKOV NETWORKS Label 1 Label 2 Label 3 113 Figure 8. it specifies the cliques and potentials between attributes of related entities at a template level. A relational clique template specifies tuples of variables in the instantiation by using a relational query language. for each link. as the unconditional case is a special case where the set of content attributes is empty. as in Fig.2. We can capture this correlation between labels by introducing. The potential on the clique will have higher values for assignments that give a common label to the linked pages. 8. 8. To specify what cliques should be constructed in an instantiation. suppose that pages with the same label tend to link to each other. doc2.2 Relational Markov networks We now extend the framework of Markov networks to the relational setting. (We provide the definitions directly for the conditional case.1: An unrolled Markov net over linked documents. The links follow a common pattern: documents with the same label tend to link to each other more often. we can write the template as a kind of SQL query: SELECT doc1. For our link example.Category.1. we will define a notion of a relational clique template. For example. A relational Markov network (RMN) specifies a conditional distribution over all of the labels of all of the entities in an instantiation given the relational structure and the content attributes. a clique between the labels of the source and the target page.Category .) Roughly speaking. so a single model provides a coherent distribution for any collection of instances from the schema.8.

W(F. F consists of doc1.r)}. This pattern can be expressed by the following template: SELECT doc1.Category.114 CHAPTER 8.Key Note the three clauses that define a query: the FROM clause specifies the cross product of entities to be filtered by the WHERE clause and the SELECT clause picks out the attributes of interest.To = doc2. doc2. A relational clique template C = (F. Doc doc2.Category.r) ensures that the entities are related to each other in specified ways. Note that the clique template does not specify the nature of the interaction between the attributes.R) is link.Rl . . the clause W(f . A clique template specifies a set of cliques in an instantiation I: C(I) ≡ {c = f .Key and F. This definition of a clique template is very flexible. and finally. ◦ W(F. Our definition of clique templates contains the corresponding three parts.Rj = Fk . where an entity variable Fi is of type E(Fi ). It allows modeling complex relational patterns on the instantiation graphs.R) — a boolean formula using conditions of the form Fi . S) consists of three components: ◦ F = {Fi } — a set of entity variables.Key and link.F rom = doc1. W. RELATIONAL MARKOV NETWORKS FROM Doc doc1. For the clique template corresponding to the SQL query above.S : f ∈ I(F) ∧ W(f .X ∪ F. To continue our webpage example. Link link WHERE link. Doc and Link.S is doc1. which will be associated with the template.Category and doc2. respectively. ◦ F. . I(F) = I(E(F1 )) × .Key ∧ link. × I(E(Fn )) denotes the cross-product of entities in the instantiation.S ⊆ F. where f is a tuple of entities {fi } in which each fi is of type E(Fi ).S selects the appropriate attributes of the entities. that is determined by the clique potentials.From = doc1.Y — a selected subset of content and label attributes in F. doc2 and link of types Doc.T o = doc2. f . as the WHERE clause of a template can be an arbitrary predicate.Category . consider another common pattern in hypertext: links in a webpage tend to point to pages of the same category.

r) 1 = Z(I. Since the clique templates do not explicitly depend on the identities of entities. I.x. I.x. I. I.x. I.Key Depending on the expressive power of our template definition language. I.r) is the normalizing partition function: Z(I. I.x. I. Link link1.r) = I. I.Key = doc2. and f is the vector .yc ) is the sum over all appearances of the template C(I) in the instantiation. Φ) specifies a set of clique templates C and corresponding potentials Φ = {φC }C∈C to define a conditional distribution: P (I.From and link1.x.r) = w f (I.r) C∈C φC (I. I.To = doc2.2. Doc doc2.xc .Key and not doc1.x.r) − log Z(I.y.Key and link2. RELATIONAL MARKOV NETWORKS 115 FROM Doc doc1. we can write log P (I. thereby allowing generalization over a wide range of different structures.To = doc1.y C∈C c∈C(I) φC (I. We can easily represent patterns involving three (or more) interconnected documents without worrying about the acyclicity constraint imposed by directed models.r) where fC (I.yc ) . Link link2 WHERE link1.From = link2. I.y | I. The RMN allows us to associate the same clique potential parameters with all of the subgraphs satisfying the template.8.yc ) c∈C(I) where Z(I.y | I. A Relational Markov network (RMN) M = (C.x. Using the log-linear representation of potentials.r) = c∈C(I) fC (I. I.xc .xc .y. we may be able to construct very complex templates that select entire subgraph structures of an instantiation. I.x. the same template can select subgraphs whose structure is fairly different. φC (VC ) = exp{wC fC (VC )}.

Maximum likelihood estimation For maximum likelihood learning. 2000] provides some theoretical justification.116 CHAPTER 8.. not just the most likely assignment.4 show. Belief Propagation (BP) is a local message passing algorithm introduced by Pearl [1988]. 1999] show that it often converges in general networks. 1999] for more details. we assume a pairwise network where all potentials are associated only with nodes and edges given by: . Empirical results [Murphy et al. For simplicity. referring to [Murphy et al. RELATIONAL MARKOV NETWORKS of all fC .3 Approximate inference and learning Applying both maximum likelihood and maximum margin learning in the relational setting is requires inference in very large and complicated networks. the RMN M produces an unrolled Markov network over the attributes of entities in I. this approach works well in our domain. there is an edge between the labels of these pages.1 illustrates a simple instance of this unrolled Markov network. There is a wide variety of approximation schemes for this problem. where exact inference is typically intractable. for a detailed description of the BP algorithm. In our webpage example. Fig. recent analysis [Yedidia et al. for every link between two pages. 8. The cliques in the unrolled network are determined by the clique templates C. We refer the reader to Yedidia et al.. the marginals are a good approximation to the correct posteriors. However. and when it does. and all of these cliques are associated with the same clique potential φC . 8. We chose to use belief propagation for its simplicity and relative efficiency and accuracy. We have one clique for each c ∈ C(I).. we need to compute basis function expectations. including MCMC and variational methods. We therefore resort to approximate methods. It is guaranteed to converge to the correct marginal probabilities for each node only for singly connected Markov networks. We provide a brief outline of one variant of BP. Given a particular instantiation I of the schema. 8. an RMN with the link basis function described above would define a Markov net in which. As our results in Sec.

Yj )ψi (yi ) yi k∈N (i)−j mki (Yi ) to compute approximate posterior “max”-marginals and use those for prediction. At each iteration. In our experiments. Yi ). Yj )ψi (Yi )ψj (Yj ) k∈N (i)−j mki (Yi ) l∈N (j)−i mlj (Yj ) These approximate marginals are precisely what we need for the computation of the basis function expectations and performing classification. each node Yi sends the following messages to all its neighbors N (i): mij (Yj ) ← α yi ψij (yi . The belief propagation algorithm is very simple. . Computing the expected basis function expectations involves summing their expected values for each clique using the approximate marginals bi (Yi ) and bij (Yi . At any point in the algorithm. Yj )ψi (yi ) k∈N (i)−j mki (Yi ) where α is a (different) normalizing constant.8. Yj ). Yj ) = φ(xij . . the marginal distribution of any node Yi is approximated by bi (Yi ) = αψi (Yi ) k∈N (i) mki (Yi ) and the marginal distribution of a pair of nodes connected by an edge is approximated by bij (Yi . Note that we can also max − product variant of loopy BP. with mij (Yj ) ← α max ψij (yi . Yi . . . ψi (Yi ) = φ(xi .3. Yj ) ij i ψi (Yi ) where ij ranges over the edges of the network and ψij (Yi . Yj ). so we use posterior marginal prediction. This process is repeated until the messages converge. this results in less accurate classification. Similarly. Yj ) = αψij (Yi . . Yn ) = 1 Z ψij (Yi . we use maxyi bi (Yi ) at prediction time. APPROXIMATE INFERENCE AND LEARNING 117 P (Y1 .

we compromised by eliminating all other pages with fewer than three outlinks.118 CHAPTER 8.. we learn a model from three schools and test the performance of the learned model on the remaining school. Texas. which we rounded independently for each label.2. 8. For each page. representing the type of webpage which is one of course. 7. we have access to the entire html of the page and the links to other pages. faculty (148). using commercial Ilog CPLEX software to solve it. Our goal is to collectively classify webpages into one of these five categories. but in a relational classification setting. research-project (82) and student (542).4 Experiments We tried out our framework on the WebKB dataset [Craven et al. Hence. 1998]. In all of our experiments. we used the LP in Sec. Texas (574). RELATIONAL MARKOV NETWORKS Maximum margin estimation For maximum margin estimation. Washington (315) and Wisconsin (454). we used the untriangulated LP we described in Sec. The resulting category distribution is: course (237). which is an instance of our hypertext example. The data set contains webpages from four different Computer Science departments: Cornell. For networks with general potentials. Washington and Wisconsin. other (332). The number of links for each school are: Cornell (574). thus evaluating the generalization performance of . Each page has a label attribute. making the number of other pages commensurate with the other categories. we used approximate LP inference inside the max-margin QP. The number of remaining pages for each school are: Cornell (280). The number of pages classified as other is quite large.4. faculty. so that a baseline algorithm that simply always selected other as the label would get an average accuracy of 75%. project or other. Texas (292). which always produced integral solutions on test data. 5. the deleted webpages might be useful in terms of their interactions with other webpages. The data set is problematic in that the category other is a grab-bag of pages of many different types. For networks with attractive potentials (AMNs). We could restrict attention to just the pages with the four other labels. The untriangulated LP produced fractional solutions for inference on the test data in several settings. student. Washington (728) and Wisconsin (1614).

Logistic Regression. but we can take additional advantage of the correlation in labels of related pages by classifying them collectively. The average error over the 4 schools was reduced by around 4% by introducing the meta-data attributes.. . Unfortunately. We found that stemming and feature selection did not provide much benefit and simply pruned words that appeared in fewer than three documents in each of the three schools in the training data. Note that meta-data. are easily incorporated as features. The results. shown in Fig. The text of the webpage is represented using a set of binary attributes that indicate the presence of different words on the page. we cannot directly compare our accuracy results with previous work because different papers use different subsets of the data and different training/test splits. in anchors of links to the page and in the last header before a link to the page [Yang et al. 1999].4. Logistic and Svm give very similar results.e. We used C ∈ [0. Incorporating meta-data gives a significant improvement. and Support Vector Machines.8.2. show that the two discriminative approaches outperform Naive Bayes. We want to capture these correlations in our model and use them for transmitting information between linked pages to provide more accurate classification. and logistic regression (Logistic) — using words and meta-words. 8. although mostly originating from pages linking into the considered page. we compare to standard text classifiers such as Naive Bayes.1. However. EXPERIMENTS 119 the different models.1 Flat models The simplest approach we tried predicts the categories based on just the text content on the webpage. 8. which have been demonstrated to be successful on this data set [Joachims. i. 2002]. linear support vector machines (Svm). 10] and took the best setting for all models. We also experimented with incorporating meta-data: words appearing in the title of the page. We experimented with several relational models.4. Our first experimental setup compares three well-known text classifiers — Naive Bayes. the resulting classification task is still flat feature-based classification.

shown in Fig. Svm. and Logistic on WebKB. Concretely. Then a simple generative model specifies a probability distribution over the existence of links between pages conditioned on both pages’ labels. we learn the probability of existence of a link between two pages given their labels. These correlations are very common in our data: courses and research projects almost never link to each other. faculty rarely link to each other. Note that this model does not require inference during learning. For each page.15 0.1 0.We train this model using maximum conditional likelihood (labels given the words and the links) and maximum margin. a logistic regression model predicts the page label given the words and meta-features. Maximum likelihood estimation (with regularization) of the generative component is closed .) 8.2 0.2: Comparison of Naive Bayes.2 Link model Our first model captures direct correlations between labels of linked pages.05 0 Words Words+Meta Naïve Bayes Svm Logistic Figure 8. captures this correlation through links: in addition to the local bag of words and meta-data attributes. The Exists-ML model is a (partially) generative model proposed by Getoor et al. 8.3 0.25 Test Error 0. with and without meta-data features. we introduce a relational clique template over the labels of two pages that are linked.1. We also compare to a directed graphical model to contrast discriminative and generative models of relational structure.120 CHAPTER 8. The Link model.4. [2001]. (Only averages over the 4 schools are shown here. RELATIONAL MARKOV NETWORKS 0. students have links to all categories but mostly courses.35 0.

3: Comparison of flat versus collective classification on WebKB: SVM. but also those that do not exist. since the resulting graphical model includes edges not only for the existing hyperlinks. The Exists model doesn’t perform very well in comparison. the prediction phase is much more expensive.3 shows a gain in accuracy from SVMs to the Link model by using the correlations between labels of linked web pages.4. 8. Exists model with logistic regression and the Link model estimated using the maximum likelihood (ML) and the maximum margin (MM) criteria.4. form given appropriate co-occurrence counts of linked pages’ labels. There is also very significant additional gain by using maximum margin training: the error rate of Link-MM is 40% lower than that of Link-ML. EXPERIMENTS 121 30% 25% 20% Exists-ML SVM Link-ML Link-MM Error 15% 10% 5% 0% Cor Tex Was Wis Average Figure 8. Intuitively. By contrast. the Link model avoids this problem by only modeling the conditional distribution given the existing links. 8. Fig.8. This can be attributed to the simplicity of the generative model and the difficulty of the resulting inference problem.3 Cocite model The second relational model uses the insight that a webpage often has internal structure that allows it to be broken up into sections. linked or not. For example. and 51% lower than multi-class SVMs. a faculty webpage might have . However. observing the link structure directly correlates all page labels in a website.

we have a relational clique template defined by: SELECT doc1. Intuitively.Section = link2. and a third to his advisees. To take advantage of this trend. We can view a section of a webpage as a fine-grained version of Kleinberg’s hub [Kleinberg. and Logistic on WebKB.From = link2.To = doc2.Key We compared the performance of SVM.Category FROM Doc doc1. Doc doc2. with a list of links to all of the projects of the faculty member. Note that we expect the correlation between the labels in this case to be positive. RELATIONAL MARKOV NETWORKS 25% 20% SVM Cocite-ML Cocite-MM Error 15% 10% 5% 0% Cor Tex Was Wis Average Figure 8.Key and not doc1. so we can use AMN-type potentials in the max-margin estimation.To = doc1. a second section might contain links to the courses taught by the faculty member.) one section that discusses research. Svm. 1999] (a page that contains a lot of links to pages of particular category). Link link1. The results. Link link2 WHERE link1. or linked to from the same section. Cocite-ML and Cocite-MM. (Only averages over the 4 schools are shown here. In the RMN.122 CHAPTER 8. if two pages are cocited. they are likely to be on similar topics. we need to enrich our schema by adding the attribute Section to Link to refer to the section number it appears in. The Cocite model captures this type of correlation.Key and link2. .From and link1.Category.4: Comparison of Naive Bayes. doc2. with and without meta-data features.Key = doc2.Section and link1. We defined a section as a sequence of three or more links that have the same path to the root in the html parse tree.

where the presence of an A-B link and of a B-C link increases (or decreases) the likelihood of an A-C link. a professor X and two other entities Y and Z.. Later work showed how to efficiently learn model parameters and structure (equivalent of clique selection in Markov networks) from data [Friedman et al. if Y is Professor X’s advisee. which suggests that the optimization successfully avoided problematic parameterizations of the network. even in the case of the non-optimal multi-class relaxation. Our framework accommodates these patterns easily. If X’s webpage mentions Y and Z in the same context. we can capture such patterns as well. We can incorporate even more complicated patterns. in our experiments. 8. Again. [2002] propose several generative models of relational structure.5 Related work Our RMN representation is most closely related to the work on PRMs [Koller & Pfeffer. also demonstrate significant improvements of the relational models over the SVM. it is likely that the X-Y relation and the Y-Z relation are of the same type. where objects that share a certain graph-based property are more likely to have the same label. which shows a 30% relative reduction in test error over SVM. We give some examples here. then probably so is Z.4. By introducing cliques over triples of relations. 1999]. the learned Cocite-MM weights never produced fractional solutions when used for inference.8. Their approach easily captures the dependence of link existence on attributes of entities. Getoor et al. Consider. in particular those that involve several links at a time. One useful type of pattern type is a similarity template. Another useful type of subgraph template involves transitivity patterns. However there are many patterns that we are difficult to model in PRMs. For example. We note that. but of course we are limited by the ability of belief propagation .5. 1998]. 8. maximum likelihood trained model Cocite-ML achieves a worse test error than maximum margin Cocite-MM model. students often assist in courses taught by their advisor. by introducing pairwise cliques between the appropriate relation variables. Note that this type of interaction cannot be accounted for just using pairwise cliques. for example. RELATED WORK 123 shown in Fig. for example. The improvement is present when testing on each of the schools.

it is possible to simply view the presence of particular citations as atomic features. In our WebKB example. Chakrabarti et al. In some cases.6 Conclusion In this chapter. 1990] and hypertext [Kleinberg. for the transitivity pattern. with no formal justification. it is not clear how to parameterize a link existence variable for a link that is involved in multiple triangles. However. we might consider simply directing the correlation edges between link existence variables arbitrarily. 1999]. Our approach provides a coherent foundation for the process of collective classification. none of these approaches provide a coherent model for the correlations between linked webpages. we can incorporate relational features into standard flat classification. this approach is limited in cases where some or even all of the relational features that occur in the test data are not observed in the training data. achieving a significant improvement compared to a baseline of using the text in each document alone. The structure of the relational graph has been used extensively to infer importance in scientific publications [Egghe & Rousseau. 8. there is no overlap between the webpages in the . applying combinations of classifiers in a procedural way. when classifying papers into topics. exploiting the interactions between their labels.124 CHAPTER 8. However. However. Several recent papers have proposed algorithms that use the link graph to aid classification. A similar approach was used by Neville and Jensen [2000] in a different domain. We have shown that we can exploit a very rich set of relational patterns in classification. [1998] use system-predicted labels of linked documents to iteratively re-label each document in the test set. 2003b]. RELATIONAL MARKOV NETWORKS to scale up as we introduce larger cliques and tighter loops in the Markov network. Attempts to model such pattern in PRMs run into the constraint that the probabilistic dependency graph (Bayesian network) must be a directed acyclic graph. where we want to classify multiple entities. We describe and exploit these patterns in our work on RMNs using maximum likelihood estimation [Taskar et al. significantly improving the classification accuracy over standard flat classification. we propose a new approach for classification in relational domains. and used these pages to improve classification of university webpages. For example. Slattery and Mitchell [2000] tried to identify directory (or hub) pages that commonly list pages of the same topic.. For example.

social networks provide extensive information about interactions among people and organizations. words) from the related entities can aid in classification. 2003b]). RMNs offer a principled method for learning to predict communities of and hierarchical structure between people and organizations based on both the local attributes and the patterns of static and dynamic interaction. RMNs are generally applicable to any relational domain. CONCLUSION 125 different schools.. it is particularly interesting to explore the problem of inducing them automatically. The results in this chapter represent only a subset of the domains we have worked on (see [Taskar et al. so we cannot learn anything from the training data about the significance of a hyperlink to/from a particular webpage in the test data.g. Given the wealth of possible patterns. . In particular.8. however. but cannot exploit the strong correlation between the labels of related entities that RMNs capture. Incorporating basic features (e. Hypertext is the most easily available source of structured data.6..

126 CHAPTER 8. RELATIONAL MARKOV NETWORKS .

Part III Broader applications: parsing. clustering 127 . matching.

u For clarity of presentation.Chapter 9 Context free grammars We present a novel discriminative approach to parsing using structured max-margin criterion based on the decomposition properties of context free grammars. Terminal symbols for our purposes are part-of-speech tags like nouns (NN). we restrict our grammars to be in Chomsky normal form as in Sec. For example. 9.1(a) shows a parse tree for the sentence The screen was a sea of 128 . We show experimental evidence of the model’s improved performance over a natural baseline model and a lexicalized probabilistic context-free grammar. The non-terminal symbols correspond to syntactic categories such as noun phrase (NP) or verbal phrase (VP). Fig. we are not concerned with defining a distribution over sequences of words (language model). verbs (VBD). determiners (DT). Our models can condition on arbitrary features of input sentences.4. We show that this framework allows high-accuracy parsing in cubic time by exploiting novel kinds of lexical information. although many others have also been proposed [Manning & Sch¨ tze. However. 9. The terminal symbols are usually words of the sentence. in the discriminative framework that we adopt. we condition on the words in a sentence to produce a model of the syntactic structure. 1999]. thus incorporating an important kind of lexical information not usually used by conventional parsers.1 Context free grammar model CFGs are one of the primary formalisms for capturing the recursive structure of syntactic constructions. 3. Instead.

n is constrained to be in NS . PU = {A → D : A ∈ N . 1993].9.1. P = {PB . A CFG defines a set of valid parse trees in a natural manner: Definition 9. 3. NN.4 here: Definition 9. PB = {A → B C : A.. we repeat our definition of a CFG from Sec. B.e taking values in N ∪⊥ to represent the label of the subtree that exactly covers. C ∈ N } and Unary productions. For each span starting with s and an ending with e. This tree representation seems quite different from the graphical models we have been considering thus far.1. respectively. B and C. PU }. where the set of valid labels of the internal nodes other than the root is N and the set of valid labels for the leaves is T . However. we can use an equivalent representation that essentially encodes a tree as an assignment to a set of appropriate variables. or dominates. CONTEXT FREE GRAMMAR MODEL 129 red. the words of the sentence from s to e. All other internal nodes have two children. there are exponentially many parse trees that produce a sentence of length n.1 (CFG) A CFG G consists of: ◦ A set of non-terminal symbols. In general.1. Additionally. hence 0 ≤ s < e ≤ n for a sentence of length n. D ∈ T }. For convenience. The value ⊥ is assigned if no subtree dominates the span. DT. each pre-leaf node has a single child and this pair of nodes can be labeled as A and D. N ◦ A designated set of start symbols. T ◦ A set of productions. rather than to words themselves. and this triple of nodes can be labeled as A. The “top” symbol Y0. if and only if there is a unary production A → D ∈ PU . The non-terminal symbols with bars (for example. The set of symbols we use is based on the Penn Treebank [Marcus et al. The root’s label set is NS . Indices s and e refer to positions between words. if and only if there is a binary production A → B C ∈ PB . divided into Binary productions. respectively. VBD) are added to conform to the CNF restrictions. we introduce a variable Ys.2 (CFG tree) A CFG tree is a labeled directed tree. since it represents . NS ⊆ N ◦ A set of terminal symbols. left and right.

e . For example.e = ⊥.s taking values in T to represent the terminal symbol (part-of-speech) between s and s + 1. The row and column number in the grid correspond to the beginning and end of the span. The gray squares on the diagonal represent the terminal variables. Empty squares correspond to ⊥ values. We also introduce variables Ys. 9. While any parse tree corresponds to an assignment to this set of variables in a straightforward manner. 9. it is often called a constituent.130 CHAPTER 9.1(b) shows the representation of the tree in Fig. CONTEXT FREE GRAMMARS (a) (b) Figure 9. the starting symbol of the sentence.1: Two representations of a binary parse tree: (a) nested tree structure. The gray squares on the diagonal represent part-of-speech tags. The row and column number are the beginning and end of the span. the converse is not true: there are assignments that do not correspond to valid parse trees.0 = DT.1(a) as a grid where each square corresponds to Ys.6 = NN.5 = NP and Y1. and (b) grid of labeled spans. consider the . Y3. If Ys. respectively. Y6. In order to characterize the set of valid assignments Y. respectively. Empty squares correspond to non-constituent spans.4 = ⊥. Fig. the figure shows Y0.

m and ym . will produce the rest of the tree down to the pre-terminals.e = A ∧ ys.2) holds because that if the span from s to e is dominated by a subtree starting with A (ys. ys. (9. The second set of constraints (9. I(y e<e ≤n 0 ≤ s < e ≤ n.e = (B. C. The second set of constraints (9.1-9. In fact the converse is true as well: Theorem 9. Similarly. D) abbreviates ys.s = D) = I(y 1 s.n .2) B→C A∈PB 1 s. Starting from the root symbol.3) (9. then there must be a unique production starting with A and some split point m. s < m < e. C)). ∀A ∈ N .1) holds because if the span from s to e is dominated by a subtree starting with A (that is. C) abbreviates ys. D)).m and ym. 0 ≤ s < n.e = A). y0.s. Conversely. B. C)) I(y 0≤s <s 0 ≤ s < e ≤ n. 0 ≤ s < n.n . the first set of constraints (9.s.e = C and ys.1) ensures that a unique production spans 0 to n. A.e = A.m = B ∧ ym.s+1 = A) = I(y 1 s. ∀A ∈ N .n .m.e = A).m. y0. that produces that subtree. I(y 1 s. I(y (9.1. then there must be a (unique) production that generated it: either starting before s or after e.1) A→B C∈PB 1 s. Proof sketch: It is straightforward to construct a parse tree from y ∈ Y in a top-down manner. The last two sets of constraints (9. D)).4) .1. I(y 1 s . ∀A ∈ N .n . CONTEXT FREE GRAMMAR MODEL 131 constraints that hold for a valid assignment y: 1 s.2) ensures that all other spans y0. if ys.e = (B.s+1 = A ∧ ys.3 and 9.e = (A. B.4) hold since the terminals are generated using valid unary productions.s+1 = (A.e = (A. Recursing on the two subtrees. say splitting at m and specifying the values for y0. ∀D ∈ T .s = D. then y represents a valid CFG tree.s.s+1 = (A.4) as Y. We denote the set of assignments y satisfying (9.3 If y ∈ Y.m and ym.e = A) = I(y + B→A C∈PB 1 s.e. A)). (9. no productions start with A and cover s to e.4) A→D∈PU A→D∈PU The notation ys. the third and fourth set of constraints (9.s+1 = (A.s.e = A) = I(y s<m<e 1 s.9.3 and 9. for m = m are labeled by ⊥. The first set of constraints (9.

u The algorithm computes the highest score of any subtree starting with a symbol over each span 0 ≤ s < e ≤ n recursively: ∗ Ss. (9.e ) + 0≤s<n Ss.5). 1999]. m. ∀A ∈ N . 0 ≤ s < n.m. B.m.m.n (A).e (C). More generally. This can be done simply by assigning a probability to each production and making sure that the sum of probabilities of all productions starting with each symbol is 1: P (A → B C) = 1.s+1 (A) = ∗ Ss.s+1 ).132 CHAPTER 9. C) + Ss. a weighted CFG assigns a score to each production (this score may depend on the position of the production s.m.m (B) + Sm.s+1 (A. ∀A ∈ N . ∗ ∗ Ss.m = ⊥ ∨ ym. then the tree score is the tree log probability. If the production scores are production log probabilities.e = ⊥).e ) = 0 if (ys. weighted CFGs do not have the local normalization constraints Eq. CONTEXT FREE GRAMMARS ensure that the terminals are generated by an appropriate unary productions from the preterminals. 9.C:A→B C∈PB D:A→D∈PU P (A → D) = 1.2 Context free parsing A standard approach to parsing is to use a CFG to define a probability distribution over parse trees.5) 0 ≤ s < e ≤ n. B.s+1 (ys. Using the arg max’s of the max’s in .e (A. e) such that the total score of a tree is the sum of the score of all the productions used: S(y) = 0≤s<m<e≤n Ss.e = ⊥ ∨ ys. D).e (ys. Manning & Sch¨ tze. However.e (ys. ∀A ∈ N .m. (9. A→B C∈PB max s<m<e ∗ The highest scoring tree has score maxA∈NS S0. We can use a Viterbi-style dynamic programming algorithm called CKY to compute the highest score parse tree in O(|P|n3 ) time [Younger. 1967. where Ss.e (A) = A→D∈PU max Ss. The probability of a tree is simply the product of probabilities of the productions used in the tree.

which includes both spans and span triplets: C = {(s.e . e) : 0 ≤ s ≤ m < e ≤ n}.e ) = 1 s. where n is the length of the sentence x and xs. NP.m. y) = 0≤s≤e≤n f (xs.e ) + 0≤s≤m<e≤n f (xs. where the basis functions consist of the counts of the productions used in the parse. ys. We assume that score ties are broken in a predetermined way. VP) ∧ I(y mth word(x) = was)). where X is a set of sentences.e ).3. f (xs.3 Discriminative parsing models We cast parsing as a structured classification task. y).e . m. we introduce the set of indices. e. C.e = S. Note that this class of discriminants includes PCFG models. say according to some lexicographic order of the symbols. yc ). we can back-trace the highest scoring tree itself.e are the relevant subsets of the sentence the basis functions depend on.m. The functions we consider take the following linear discriminant form: hw (x) = arg max w f (x. and the parameters w are the log-probabilities of those productions.e . ys. We assume that the basis functions decompose with the CFG structure: f (x. and neighboring positions in the sentence x (e. 9.m. where we want to learn a function h : X → Y.m. Hence. m. y) = c∈C f (xc . y where w ∈ IRd and f is a basis function representation of a sentence and parse tree pair f : X × Y → IRd . m) : 0 ≤ s ≤ e ≤ n} ∪ {(s.m. For example. f could include functions which identify the production used together with features of the words at positions s.g.e and xs. We could also include functions that identify the label of the span . To simplify notation. DISCRIMINATIVE PARSING MODELS 133 the computation of S ∗ .m.9. and Y is a set of valid parse trees according to a fixed CFG grammar. f (x. ys.

134

CHAPTER 9. CONTEXT FREE GRAMMARS

from s to e together with features of the word (e.g. f (xs,m , ys,m ) = 1 s,m = NP) ∧ I(y sth word(x) = the)).

9.3.1 Maximum likelihood estimation
The traditional method of estimating the parameters of PCFGs assumes a generative model that defines P (x, y) by assigning normalized probabilities to CFG productions. We then maximize the joint log-likelihood
i

log P (x(i) , y(i) ) (with some regularization). We com-

pare to such a generative grammar of Collins [1999] in our experiments. A alternative probabilistic approach is to estimate the parameters discriminatively by maximizing conditional log-likelihood. For example, the maximum entropy approach [Johnson, 2001] defines a conditional log-linear model: Pw (y | x) = where Zw (x) = the sample,
i

1 exp{w f (x, y)}, Zw (x)

y

exp{w f (x, y)}, and maximizes the conditional log-likelihood of

log P (y(i) | y(i) ), (with some regularization). The same assumption that

the basis functions decompose as sums of local functions over spans and productions is typically made in such models. Hence, as in Markov networks, the gradient depends on the expectations of the basis functions, which can be computed in O(|P|n3 ) time by dynamic programming algorithm called inside-outside, which is similar to the CKY algorithm. However, computing the expectations over trees is actually more expensive in practice than finding the best tree for several reasons. CKY works entirely in the log-space, while inside-outside needs to compute actual probabilities. Branch-and-prune techniques, which save a lot of useless computation, are only applicable in CKY. A typical method for finding the parameters is to use Conjugate Gradients or L-BFGS methods [Nocedal & Wright, 1999; Boyd & Vandenberghe, 2004], which repeatedly compute these expectations to calculate the gradient. Clark and Curran [2004] report experiments involving 479 iterations of training for one model, and 1550 iterations for another using similar methods.

9.3. DISCRIMINATIVE PARSING MODELS

135

9.3.2

Maximum margin estimation

We assume that loss function also decomposes with the CFG structure: ˆ (x, y, y) = (xs,e , ys,e , ys,e ) + ˆ ˆ (xs,m,e , ys,m,e , ys,m,e ) =
c∈C

ˆ (xc , yc , yc ).

0≤s≤e≤n

0≤s≤m<e≤n

One approach would be to define (xs,e , ys,e , ys,e ) = 1 s,e = ys,e ). This would lead to ˆ I(y ˆ ˆ ˆ (x, y, y) tracking the number of “constituent errors” in y. Another, more strict definition ˆ ˆ would be to define (xs,m,e , ys,m,e , ys,m,e ) = 1 s,m,e = ys,m,e ). This definition would lead I(y ˆ ˆ to (x, y, y) being the number of productions in y which are not seen in y. The constituent loss function does not exactly correspond to the standard scoring metrics, such as F1 or crossing brackets, but shares the sensitivity to the number of differences between trees. We have not thoroughly investigated the exact interplay between the various loss choices and the various parsing metrics. We used the constituent loss in our experiments. As in the max-margin estimation for Markov networks, we can formulate an exponential size QP: min s.t. 1 ||w||2 + C 2 ξi
i

(9.6)

w ∆fi (y) ≥ i (y) − ξi ∀i, y,

where ∆fi (y) = f (x(i) , y(i) ) − f (x(i) , y), and i (y) = (x(i) , y(i) , y). The dual of Eq. (9.6) (after normalizing by C) is given by: max
i,y

1 αi (y) i (y) − C 2 αi (y) = 1,

2

αi (y)∆fi (y)
i,y

(9.7)

s.t.
y

∀i; αi (y) ≥ 0, ∀i, y.

Both of the above formulations are exponential (in the number of variables or constraints) in the lengths (ni ’s) of the sentences. But we can exploit the context-free structure of the basis functions and the loss to define a polynomial-size dual formulation in terms of

136

CHAPTER 9. CONTEXT FREE GRAMMARS

marginal variables µi (y): µi,s,e (A) ≡
y:ys,e =A

αi (y), 0 ≤ s ≤ e ≤ n, ∀A ∈ N ; αi (y),
y:ys,s =D

µi,s,s (D) ≡ µi,s,m,e (A, B, C) ≡

0 ≤ s ≤ e ≤ n, ∀D ∈ T ; αi (y), 0 ≤ s < m < e ≤ n, ∀A → B C ∈ PB , 0 ≤ s < n, ∀A → D ∈ PU .

y:ys,m,e =(A,B,C)

µi,s,s,s+1 (A, D) ≡
y:ys,s,s+1 =(A,D)

αi (y),

There are O(|PB |n3 + |PU |ni ) such variables for each sentence of length ni , instead of i exponentially many αi variables. We can now express the objective function in terms of the marginals. Using these variables, the first set of terms in the objective becomes: αi (y) i (y) =
i,y i,y

αi (y)
c∈C (i)

i,c (yc )

=
i,c∈C (i) ,yc

µi,c (yc )

i,c (yc ).

Similarly, the second set of terms (inside the 2-norm) becomes: αi (y)∆fi (y) =
i,y i,y

αi (y)
c∈C (i)

∆fi,c (yc ) =
i,c∈C (i) ,yc

µi,c (yc )∆fi,c (yc ).

As in M3 Ns, we must characterize the set of marginals µ that corresponds to valid α. The constraints on µ are essentially based on the those that define Y in (9.1-9.4). In addition, we require that the marginals over the root nodes, µi,0,ni (y0,ni ), sums to 1 over the possible start symbols NS .

9.3. DISCRIMINATIVE PARSING MODELS

137

Putting the pieces together, the factored dual is:
2

max
i,c∈C (i)

µi,c (yc )

i,c (yc )

+C
i,c∈C (i)

µi,c (yc )∆fi,c (yc ) µi,c (yc ) ≥ 0; ∀i, ∀c ∈ C (i) ; ∀i, 0 ≤ s < e ≤ ni , ∀A ∈ N ;

(9.8)

s.t.
A∈NS

µi,0,ni (A) = 1, µi,s,e (A) =
s<m<e

∀i;

µi,s,m,e (A, B, C), µi,s ,s,e (B, A, C)
0≤s <s

A→B C∈PB

µi,s,e (A) = + µi,s,s+1 (A) = µi,s,s (D) =

B→A C∈PB

µi,s,e,e (B, C, A), ∀i, 0 ≤ s < e ≤ ni , ∀A ∈ N ; µi,s,s,s+1 (A, D), µi,s,s,s+1 (A, D), ∀i, 0 ≤ s < ni , ∀A ∈ N ; ∀i, 0 ≤ s < ni , ∀D ∈ T .

B→C A∈PB e<e ≤ni

A→D∈PU

A→D∈PU

The constraints on µ is necessary, since they must correspond to marginals of a distribution over trees. They are also sufficient: Theorem 9.3.1 A set of marginals µi (y) satisfying the constraints in Eq. (9.8) corresponds to a valid distribution over the legal parse trees y ∈ Y (i) . A consistent distribution αi (y) is given by αi (y) = µi,0,ni (y0,ni )
0≤s≤m<e≤ni

µi,s,m,e (ys,m,e ) , µi,s,e (ys,e )

where 0/0 = 0 by convention. Proof sketch: The proof follows from inside-outside probability relations [Manning & Sch¨ tze, 1999]. The first term is a valid distribution of starting symbols. Each u
µi,s,m,e (ys,m,e ) µi,s,e (ys,e )

term for m > s corresponds to a conditional distribution over binary productions (ys,e → ys,m ym,e ) that are guaranteed to sum to 1 over split points m and possible productions. Similarly, each
µi,s,s,s+1 (ys,s,s+1 ) µi,s,s+1 (ys,s+1 )

term for corresponds to a conditional distribution over

138

CHAPTER 9. CONTEXT FREE GRAMMARS

unary productions (ys,s+1 → ys,s ) that are guaranteed to sum to 1 over possible productions. Hence, we have defined a kind of PCFG (where production probabilities depend on the location of the symbol), which induces a valid distribution αi over trees. It straightforward to verify that this distribution has marginals µi .

9.4 Structured SMO for CFGs
We trained our max-margin models using the Structured SMO algorithm with block-coordinate descent adopted from graphical models (see Sec. 6.1). The CKY algorithm computes similar max-marginals in the course of computing the best tree as does Viterbi in Markov networks. vi,c (yc ) = max [w fi (y) + i (y)],
y∼yc

αi,c (yc ) = max αi (y).
y∼yc

We also define vi,c (yc ) = maxyc =yc vi,c (yc ) = maxy∼yc [w fi (y) + i (y)]. Note that we do not explicitly represent αi (y), but we can reconstruct the maximum-entropy one from the marginals µi as in Theorem 9.3.1. We again express the KKT conditions in terms of the max-marginals for each span and span triple c ∈ C (i) and its values yc : αi,c (yc ) = 0 ⇒ vi,c (yc ) ≤ vi,c (yc ); αi,c (yc ) > 0 ⇒ vi,c (yc ) ≥ vi,c (yc ). (9.9)

The algorithm cycles through the training sentences, runs CKY to compute the maxmarginals and performs an SMO update on the violated constraints. We typically find that 20-40 iterations through the data are sufficient for convergence in terms of the objective function improvements.

9.5 Experiments
We used the Penn English Treebank for all of our experiments. We report results here for each model and setting trained and tested on only the sentences of length ≤ 15 words. Aside

9.5. EXPERIMENTS

139

from the length restriction, we used the standard splits: sections 2-21 for training (9753 sentences), 22 for development (603 sentences), and 23 for final testing (421 sentences). As a baseline, we trained a CNF transformation of the unlexicalized model of Klein and Manning [2003] on this data. The resulting grammar had 3975 non-terminal symbols and contained two kinds of productions: binary non-terminal rewrites and tag-word rewrites. Unary rewrites were compiled into a single compound symbol, so for example a subjectgapped sentence would have label like S + VP. These symbols were expanded back into their source unary chain before parses were evaluated. The scores for the binary rewrites were estimated using unsmoothed relative frequency estimators. The tagging rewrites were estimated with a smoothed model of P (w|t), also using the model from Klein and Manning [2003]. In particular, Table 9.2 shows the performance of this model (GENERATIVE): 87.99 F1 on the test set. For the
BASIC

max-margin model, we used exactly the same set of allowed rewrites

(and therefore the same set of candidate parses) as in the generative case, but estimated their weights using the max-margin formulation with a loss that counts the number of wrong spans. Tag-word production weights were fixed to be the log of the generative P (w|t) model. That is, the only change between GENERATIVE and BASIC is the use of the discriminative maximum-margin criterion in place of the generative maximum likelihood one for learning production weights. This change alone results in a small improvement (88.20 vs. 87.99 F1 ). On top of the basic model, we first added lexical features of each span; this gave a
LEXICAL

model. For a span s, e of a sentence x, the base lexical features were:

◦ xs , the first word in the span ◦ xs−1 , the preceding adjacent word ◦ xe−1 , the last word in the span ◦ xe , the following adjacent word ◦ xs−1 , xs ◦ xe−1 , xe ◦ xs+1 for spans of length 3

140

CHAPTER 9. CONTEXT FREE GRAMMARS

Model GENERATIVE BASIC LEXICAL LEXICAL+AUX

P 87.70 87.51 88.15 89.74

R 88.06 88.44 88.62 90.22

F1 87.88 87.98 88.39 89.98

Table 9.1: Development set results of the various models when trained and tested on Penn treebank sentences of length ≤ 15.

Model GENERATIVE BASIC LEXICAL LEXICAL+AUX COLLINS 99

P 88.25 88.08 88.55 89.14 89.18

R F1 87.73 87.99 88.31 88.20 88.34 88.44 89.10 89.12 88.20 88.69

Table 9.2: Test set results of the various models when trained and tested on Penn treebank sentences of length ≤ 15.

These base features were conjoined with the span length for spans of length 3 and below, since short spans have highly distinct behaviors (see the examples below). The features are lexical in the sense than they allow specific words and word pairs to influence the parse scores, but are distinct from traditional lexical features in several ways. First, there is no notion of headword here, nor is there any modeling of word-to-word attachment. Rather, these features pick up on lexical trends in constituent boundaries, for example the trend that in the sentence The screen was a sea of red., the (length 2) span between the word was and the word of is unlikely to be a constituent. These non-head lexical features capture a potentially very different source of constraint on tree structures than head-argument pairs, one having to do more with linear syntactic preferences than lexical selection. Regardless of the relative merit of the two kinds of information, one clear advantage of the present approach is that inference in the resulting model remains cubic (as opposed to O(n5 )), since the dynamic program need not track items with distinguished headwords. With the addition of these features, the accuracy moved past the generative baseline, to 88.44.

The auxiliary classifiers were then retrained on the entire training set. last-following. and following tag in the span.94 / F1 70.62 each. last. The second feature was the output of a flat classifier which was trained to predict whether single spans. this feature added little information. auxiliary classifiers. and using their predictions as features on the other fold. The first feature was the prediction of the generative baseline.77 / R 63. 1999]. As a result. this analysis is counter to the trend of today to be a one-word constituent. consider the sentence The Egyptian president said he would visit Libya today to resume the talks. The generative model incorrectly considers Libya today to be a base NP..12 F1 . These features represent the preference of the word today for being the first and last word in constituent spans of length 1. Another kind of feature that can usefully be incorporated into the classification process is the output of other. the resulting max-margin model (LEXICAL + AUX) scored 89.5. this model makes this parse decision correctly. Two features relevant to this trend are: (CONSTITUENT ∧ first-word = today ∧ length = 1) and (CONSTITUENT ∧ last-word = today ∧ length = 1). 2003]. but also the following: the preceding. and the entire tag sequence. first-last. its weight will be artificially (and detrimentally) high. pairs of tags such as preceding-first.which In this length 1 case.58). were constituents or not.1 In the LEXICAL model. We used two such auxiliary classifiers. Otherwise. these features have quite large positive weights: 0. Tag features on the test sets were taken from a pretagging of the sentence by the tagger described in [Toutanova et al. 1 . one must take care that its reliability on the training not be vastly greater than its reliability on the test set. not bracket labels). To situate these numbers with respect to other models. based on a bundle of features including the list above. first. we split the training into two folds. We then trained the auxiliary classifiers on each fold. but made the learning phase faster. in isolation. and their predictions used as features on the development and test sets.While the flat classifier alone was quite poor (P 78. giving a prediction feature for each span (these classifiers predicted only the presence or absence of a bracket over that span. However. To ensure that such features are as noisy on the training data as the test data. the parser in [Collins.9. EXPERIMENTS 141 As a concrete (and particularly clean) example of how these features can sway a decision. these are the same feature. Note also that the features are conjoined with only one generic label class “constituent” rather than specific constituent types. preceding-following. For this kind of feature.

Given sufficiently “local” features. an initial parser is used to generate a number of candidate parses. 2004. and intricately smoothed scores 88. 9. 1999. Miyao & Tsujii. 2001. Miyao & Tsujii. 2002. which were based on maximum entropy estimation. the decoding and parameter estimation problems can be solved using dynamic programming algorithms. 2003]. Geman & Johnson. Clark & Curran.. Moreover. Collins. CONTEXT FREE GRAMMARS is generative. our method incorporates an articulated loss function which penalizes larger tree discrepancies more severely than smaller ones. The method we presented has several compelling advantages. 41% of the correct parses were not in the candidate pool of ∼30-best parses. rather than expectations over all trees (for example using the inside-outside algorithm). 2004]. the structured SMO we use requires only the calculation of Viterbi trees.. Unlike reranking methods. Shen et al. Collins. Broadly speaking. in the work of Collins [2000]. Kaplan et al. In reranking methods [Johnson et al. where variants of the inside-outside algorithm can be used to efficiently calculate gradients of the log-likelihood function. 2002. reranking and dynamic programming approaches. 2004] are based on conditional log-linear (maximum entropy) models. 1999. despite the exponential number of trees represented by the parse forest. In dynamic programming methods. 2004. lexicalized. Collins. Unlike previous dynamic programming approaches. A discriminative model is then used to choose between these candidates. 2002. Geman & Johnson. a large number of candidate parse trees are represented compactly in a parse tree forest or chart.69 over the same train/test configuration. For example.6 Related work A number of recent papers have considered discriminative approaches for natural language parsing [Johnson et al.. several approaches [Johnson. 2000.. For example.. . Johnson. these approaches fall into two categories. This distinction can be very significant. 2001. as the set of n-best parses often does not contain the true parse. 2002. Kaplan et al. our method is an endto-end discriminative model over the full space of parses. 2000.142 CHAPTER 9. Clark & Curran. which consider only a pre-pruned selection of “good” parses. 2004.

20-40 iterations were generally required for convergence (except the iterations. CONCLUSION 143 This allows a range of optimizations that prune the space of parses (without making approximations) not possible for maximum likelihood approaches which must extract basis function expectations from the entire set of parses. since they must only collect counts from the training set. . discriminative methods are inherently expensive. It is worth considering the cost of this kind of method.) BASIC model. while still decomposing in a way that exploits the shared substructure of parse trees in the standard way. which allows a discriminative SVM-like objective to be applied to the parsing problem. In our experiments.9. On the other hand. accuracy and efficiency of a parsing model is an important area of future research.7 Conclusion We have presented a maximum-margin approach to parsing. This trade-off between the complexity. the max-margin approach does have the potential to incorporate many new kinds of features over the input. since they all involve iteratively checking current model performance on the training set. Our framework permits the use of a rich variety of input features.7. At training time. unlike other lexicalized models (including the Collins model which it outperforms in the present limited experiments). Generative approaches are vastly cheaper to train. and the current feature set allows limited lexicalization in cubic time. which took about 100 9. which means parsing the training set (usually many times).

we use the problem of disulfide connectivity prediction as an example. 4. Recently proposed models have formulated this prediction task as a maximum weight perfect matching problem in a graph containing cysteines as nodes with edge weights measuring the attraction strength of the potential bridges. find a parameterized edge scoring function such that the correct matchings have the highest score. Throughout this chapter. We motivate this problem using an application in computational biology. In this chapter. but non-bipartite matchings can be used for many other tasks. We exploit combinatorial properties of the perfect matching problem to define a compact. Bipartite matchings are used in many fields. infinite-dimensional) models and present experiments on standard protein databases. to map words of a sentence in one language to another. disulfide connectivity prediction. showing that our framework achieves state-of-the-art performance on the task. We provide some background on this problem. we focus on a more complex problem of non-bipartite matchings.Chapter 10 Matchings We address the problem of learning to match: given a set of input graphs and corresponding matchings. for example. We have shown a compact max-margin formulation for bipartite matchings in Ch. convex. 144 . Identifying disulfide bridges formed by cysteine residues is critical in determining the structure of proteins. to identify functional genetic analogues in different organisms. quadratic program. to find marker correspondences in vision problems. We use kernels to efficiently learn very rich (in-fact.

1989]. DISULFIDE CONNECTIVITY PREDICTION 145 10. Alternatively. 1994]. 2002. there are methods [Baldi et al.10. the task of predicting the connectivity pattern is typically decomposed into two subproblems: predicting the bonding state of each cysteine in the sequence.1 Disulfide connectivity prediction Proteins containing cysteine residues form intra-chain covalent bonds known as disulfide bridges. which may or may not participate in disulfide bonds. Recently. Furthermore. as well as the protein datasets. 2001]. there has been increased interest in applying computational techniques to the task of predicting the intra-chain disulfide connectivity [Fariselli & Casadio. etc.. 2004]. Thus. Fariselli et al. secondary structure or solvent accessibility information. We parameterize the this attraction strength via a linear combination of features. Such bonds are a very important feature of protein structure since they enhance conformational stability by reducing the number of configurational states and decreasing the entropic cost of folding a protein into its native state [Matsumura et al. and each edge represents the “attraction strength” between the cysteines it connects [Fariselli & Casadio. Klepeis & Floudas. Vullo & Frasconi. Since a sequence may contain any number of cysteine residues. the development of efficient. which can include the protein sequence around the two residues. Baldi et al. scalable and accurate methods for the prediction of disulfide bonds has numerous practical applications. and predicting the exact connectivity among bonded cysteines. 2001. 2004] that predict the connectivity pattern without knowing the bonding state of each cysteine1 . We thank Pierre Baldi and Jianlin Cheng for introducing us to the problem of disulfide connectivity prediction and providing us with preliminary draft of their paper and results of their model.. since the disulfide connectivity pattern imposes structure constraints. 1 .. 2003. Knowledge of the exact disulfide bonding pattern in a protein provides information about protein structure and possibly its function and evolution. We predict the connectivity pattern by finding the maximum weighted matching in a graph in which each vertex represents a cysteine residue. it can be used to reduce the search space in both protein folding prediction as well as protein 3D structure prediction. 2004.1.. They do so mostly by imposing strict structural constraints due to the resulting links between distant regions of the protein sequence [Harrison & Sternberg. evolutionary information in the form of multiple alignment profiles.

so y(i) is a binary vector of dimension Li (Li − 1)/2. then for complete graphs with even number of vertices Li . For example.2 Learning to match Formally. for example. 1ANS protein in Fig. each node is connected to at most one other node. the number of possible perfect matchings is (2ni )! 2ni ni ! (i) (which is Ω(( ni )ni ).146 CHAPTER 10. In a perfect matching. We consider a simple but very general class of attraction scoring functions defined by a weighted combination of features or basis functions: sjk (x) = d wd fd (xjk ) = w f (xjk ). We assume that the input x defines the space of possible matchings using some deterministic procedure. The space of matchings Y is very large. Let ni = Li /2. given a protein sequence. In non-perfect matchings. We represent each possible edge between nodes j and k (j < k) in example i using a binary variable yjk . X is the space of protein sequences and Y is the space of matchings of their cysteines. this model was used by Fariselli and Casadio [2001]. 15 potential bonds (edges) and 15 possible perfect matchings. but very little needs to be changed to handle sparse graphs. we seek a function h : X → Y that maps inputs x ∈ X to output matchings y ∈ Y. we assume complete graphs. For simplicity. The training data consists of m examples S = {(x(i) . super-exponential in 2 ni ). For example. y(i) )}m of input graphs and i=1 output matchings.1 has 6 cysteines (nodes). 10.1) For disulfide connectivity prediction. (10. However. (10. If example i has Li nodes. Y has interesting and complex combinatorial structure which we exploit to learn h efficiently. superexponential in the number of nodes in a graph. MATCHINGS 10. calculated by assuming that all residues in the local neighborhoods of the two cysteines make contact. in fact. Their model assigns an attraction strength sjk (x) to each pair of cysteines.2) . and summing contact potentials for pairs of residues. we construct a complete graph where each node corresponds to a cysteine. then there are Li (Li − 1)/2 edge variables. Our hypothesis class is maximum weight matchings: hs (x) = arg max y∈Y jk sjk (x)yjk . each node is connected exactly one other node.

(10.1: PDB protein 1ANS: amino acid sequence.3) Below. 10. The naive formulation of the max-margin estimation. and graph of potential disulfide bonds. y). where xjk is the portion of the input x that directly relates to nodes j and k. which enumerates all perfect 1 ||w||2 s.t. In the following sections we present two max-margin formulations. w fi (y(i) ) ≥ w fi (y) + i (y). first with an exponential set of constraints (Sec. . and then with a polynomial one (Sec. For example.4). 10. for the model: hw (x) = arg max y∈Y jk w f (xjk )yjk . the basis functions can represent arbitrary information about the two cysteine neighborhoods: the identity of the residues at specific positions around the two cysteines. y) ≡ matchings for each example i. 3D structure. we will abbreviate w f (x.10. LEARNING TO MATCH 147 RSCCPCYWGGCPWGQNCYPEGCSGPKV 1 2 3 4 5 6 2 3 1 4 5 6 Figure 10. ∀i. and w fi (y) ≡ w f (x(i) .4) The number of constraints in this formulation is super-exponential in the number of nodes in each example. or the predicted secondary structure in the neighborhood of each cysteine.3). and that our goal is to learn the weights w. Actual disulfide connectivity is shown in yellow in the 3D model and the graph of potential bonds. fd (xjk ) is a real-valued basis function and wd ∈ IR. 2 (10. We assume that the user provides the basis functions. is: min jk w f (xjk )yjk . ∀y ∈ Y (i) .2.

Under the assumption of Hamming distance loss (or any loss function that can be written as a sum of terms corresponding to edges). . The subset constraints (in the last line of Eq. these constraints are changed to µi.jk ≤ 1. where we introduce continuous variables µi. .6) µi. ∀i. Note that we have an exponential number of constraints (O(2(Li −1) )). k (i) (i) (10. j. Lawler.jk ≤ 1 require that the number of bonds incident on a node is less k or equal to one. . This problem can be solved in O(L3 ) time [Gabow. the maximum weight matching where edge jk has weight w f (xjk )+(1 −2yjk ) (plus the constant 1 y(i) ) gives the value of maxy∈Y (i) [w fi (y) + i (y)].jk = 1 to ensure exactly one bond. |V | ≥ 3 and odd. 4. but this number is asymptotically smaller than the number of possible matchings . 2 V ⊆ {1.t. Li }. 1965]. this maximization is equivalent (up to a constant term) to a maximum weighted matching problem.3 Min-max formulation Using the min-max formulation from Ch. 1 ≤ j < k ≤ Li . max jk (i) (i) (i) µi. we have a single max constraint for each i: min 1 ||w||2 s. .jk ≤ (|V | − 1). 1 ≤ j ≤ Li .6)) ensure that solutions to the LP are integral [Edmonds.148 CHAPTER 10. w fi (y(i) ) ≥ max [w fi (y) + i (y)].jk ≥ 0. Note that since the y variables are binary. 1976]. Hence. It can also be solved as a linear program.jk instead of binary variables yjk .jk [w f (xjk ) + (1 − 2yjk )] µi.t.5) The key to solving this problem efficiently is the loss-augmented inference maxy∈Y (i) [w fi (y) + i (y)]. For perfect matchings. (10. 2003]. 1973. It is an open problem to derive a polynomial sized LP formulation for perfect matchings [Schrijver. The constraints k µi. s. .k∈V 1 µi. the Hamming distance between y(i) and y can be written as (1 − y) y(i) + (1 − y(i) ) y = 1 y(i) + (1 − 2y(i) ) y. MATCHINGS 10. 2 y∈Y (i) (10.

10. ∀i. ∀i. min s.t. ∀i. The parameter C allows the user to trade off violations of the constraints with fit to the . Ai λi ≥ Fi w + ci . MIN-MAX FORMULATION 149 We can write the loss-augmented inference problem in terms of the LP in Eq. (10. Note that the dependence on w is linear and occurs only in the objective of the LP.3. Ai λi ≥ Fi w + ci . Fi is a matrix of basis function coefficients such that the component jk of the vector Fi w is w f (xjk ) and ci = (1 − 2y(i) ).6) is min λi bi s. 1 ||w||2 + C 2 (i) ξi i (10. Ai and bi are the appropriate constraint coefficient matrix and right hand side vector.5) and combine the minimization over λ with minimization over w.t.8) In case that our basis functions are not rich enough to predict the training data perfectly.t. ∀i. Ai µi ≤bi µi ≥0 y∈Y (i) where: di = 1 y(i) . The dual of the LP in Eq. (10. (10. we can introduce a slack variable ξi for each example i to allow violations of the constraints and minimize the sum of the violations: min s. λi ≥ 0. ξi ≥ 0. λi ≥ 0. (10.6): max [w fi (y) + i (y)] = di + max µi [Fi w + ci ].9) w fi (y ) + ξi ≥ di + λi bi . ∀i. µi is a vector of length Li (Li − 1)/2 indexed by bond jk. λi ≥ 0. 1 ||w||2 2 w fi (y(i) ) ≥ di + λi bi . ∀i. (10.7) (i) We plug it into Eq. respectively. Ai λi ≥ Fi w + ci . ∀i.

Theorem 10. number of bonds (x-axis) in the three formulations (perfect matching enumeration. 10. Let M be a perfect matching for a complete undirected graph G = (V. E). albeit with an exponential number of constraints. min-max and certificate). An alternating cycle is augmenting with respect to M if the score of the edges in the matching M is smaller that the score of the edges not in the matching M .1 [Edmonds.2: Log of the number of QP constraints (y-axis) vs. we develop an equivalent polynomial size formulation. In the next section.4. 1965] A perfect matching M is a maximum weight perfect matching if and only if there are no augmenting alternating cycles. . we can focus on finding a compact certificate of optimality that guarantees that y(i) = arg maxy [w fi (y) + i (y)]. Our formulation is a linearly-constrained quadratic program. We consider perfect matchings and then provide a reduction for the non-perfect case. the edges alternate between those that belong to M and those that do not.150 CHAPTER 10. MATCHINGS 30 Perfect Matchings 25 Min-max Certificate 20 Log-Number of Constraints 15 10 5 0 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Number of Bonds Figure 10. data.4 Certificate formulation Rather than solving the loss-augmented inference problem explicitly. In an alternating cycle/path in G with respect to M .

Let de . . Similarly for jk. Conj sider an alternating cycle C.10. l and the first edge. the edge (j. 0 otherwise. kj ∈ M . (10. the two shortest paths to j (one ending with M -edge and one not) contain no cycles.j+1 ≥ d1 − d0 . 2).10) Theorem 10. / ∀ jk ∈ M. Pick an arbitrary root node r. k j sjk ≥ d0 − d1 . where e = 1 if the last edge of the path is in M . j k ∀ jk ∈ M . These shortest distances are welldefined if and only if there are no negative alternating cycles. The length of the path is s(C) = s1. In the discussion below we assume that each edge score sjk has been modified this way. the shortest path to j ending with an edge not in M plus the edge jk (or kj) is an alternating path to k ending with an edge in M . with j ∈ V . j + 1) is in M .l + l−1 j=1 sj. sjk ≥ d0 − d1 .10) if and only if no augmenting alternating cycles exist. / (Only if) Suppose a distance function {de } satisfies the constraints in Eq. . . (1. Then let d0 and d1 be the length of those paths. denote the length of the shortest distance alternating path from r to j. (If) Suppose there are no augmenting alternating cycles. 1}. j e ∈ {0. 2. An alternating cycle is augmenting if and only if its length is negative. (10. for all j (for j = r. Instead. j j k k so the constraint is satisfied.10). (10. so sj. We renumber the nodes such that the cycle passes through nodes 1.4.j+1 . We begin by negating the score of those edges not in M . For each odd j. The following constraints capture this distance function. This path is longer or same length as the shortest path to k ending with an edge in M : sjk + d0 ≥ d1 (or skj + d0 ≥ d1 ). is in M .4. . Since any alternating paths from r to j can be shortened (or left the same length) by removing the cycles they contain. A condition ruling out negative length alternating cycles can be stated succinctly using a kind of distance function. We also refer to the score sjk as the length of the edge jk.2 There exists a distance function {de } satisfying the constraints in j Eq. set d0 = d1 = 0). k j sjk ≥ d1 − d0 . For j j+1 . Proof. Then for j j r r any jk (or kj) in M . so simply enumerating all of them will not do. j k sjk ≥ d1 − d0 . we can rule out such cycles by considering shortest paths. CERTIFICATE FORMULATION 151 The number of alternating cycles is exponential in the number of vertices.

min-max and certificate) is shown in Fig. (10. Finally. 1 ||w||2 2 H i w + G i di ≥ q i . we have: 1 l l−1 l−1 s(C) ≥ d0 1 − d1 l + j=1. j+1 j is not in M . (10.even Hence all alternating cycles have nonnegative length.j+1 ≥ d0 − d1 . Let di be a vector of distance variables de . The x-axis is the number of edges in the matching (number of nodes divided by two). j + 1) is not in M . In our learning formulation we have the loss-augmented edge weights sjk = (2yjk − 1)(w f (xjk )+1−2yjk ). Hi and Gi be matrices j of coefficients and qi be a vector such that Hi w + Gi di ≥ qi represents the constraints in Eq. the last edge. We extend the matching by replicating its edges in the copy and for each unmatched node. Then the following joint convex program in w and d computes the max-margin parameters: min s. in case that our basis functions are not rich enough to predict the training data perfectly. The case of non-perfect matchings can be handled by a reduction to perfect matchings as follows [Schrijver. We create a new graph by making a copy of the nodes and the edges and adding edges between each node and the corresponding node in the copy. the edge (j. Summing the edges.odd [d1 j+1 − d0 ] j + [d0 − d1 ] = 0. so s1.2. The comparison between the log-number of constraints for our three equivalent QP formulations (enumeration of all perfect matchings. We define f (xjk ) ≡ 0 for edges between the original and the copy. (i) (i) (i) Once again. MATCHINGS even j.11) ∀i.t. j+1 j j=2.10) for example i. so sj. 10. . introduce an edge to its copy. we can introduce a slack variable vector ξi to allow violations of the constraints. l). (1.152 CHAPTER 10.l ≥ d0 − d1 . 2003]. Perfect matchings in this graph projected onto the original graph correspond to non-perfect matchings in the original graph.

Let α(i) be the vector of dual variables for example i. (10. ∀i.10.12) s. (10. The only occurrence of feature vectors is in the expansion of the squared-norm term in the objective: Cyjk − αjk f (xkl ) f (x(j) ) Cyjk − αjk mn i. The primal and dual solutions are related by: w= i jk (Cyjk − αjk )f (xjk ) (i) (i) (i) (10.5 Kernels Instead of directly optimizing the primal problem in Eq.t. .j kl∈E (i) mn∈E (j) (i) (j) (i) (j) (i) (i) (i) (j) (j) (10.5. The polynomial-sized representation in Eq.11) is similarly kernelizable. (10. ∀i. Ai α(i) ≤ Cbi . and αjk is the dual variable associated with the features corresponding to the edge jk.8).13) Therefore.14) can be used to compute the attraction strength sjk (x) in a kernelized manner at prediction time.14) Eq. we can work with its dual. The dual quadratic optimization problem has the form: 2 (i) max i ci α(i) − 1 2 Cyjk − αjk f (xjk ) i jk∈E (i) (i) (i) (i) (10. we can efficiently map the original features f (xjk ) to a high-dimensional space. KERNELS 153 10. α(i) ≥ 0. Thus. we can apply the kernel trick and let f (xkl ) f (xmn ) = K(xkl . xmn ). Each training example i has Li (Li − 1)/2 dual variables.

In addition. In order to avoid biases during testing. There are 430 proteins with 2 and 3 bonds and 137 with 4 and 5 bonds. as of May 2004.154 CHAPTER 10. The DIPRO2 dataset2 was compiled and made publicly available by Baldi et al. The DIPRO2 dataset was split similarly into 5 folds.. we report results for the case where the bonding state is known. and has a total of 726 proteins. 1981] (with the BLOSUM65 scoring matrix.uci. Sequence similarity was derived using an all-against-all rigorous Smith-Waterman local pairwise alignment [Smith & Waterman. Baldi et al. 1983]. 2001]. and only 53 sequences with a higher number of bonds.ics. which contain intra-chain disulfide bonds. MATCHINGS 10. It consists of all proteins from PDB [Berman et al. [2004]. the sequences are annotated with secondary structure and solvent accessibility information derived from the DSSP database [Kabsch & Sander. with the constraint that proteins no proteins with sequence similarity of more than 30% belong to different subsets. Fariselli & Casadio. and we have followed the same procedure for extracting sequences from the database. Vullo & Frasconi. It contains only sequences with experimentally verified disulfide bridges.. Pairs of chains whose alignment is less than 30 residues were considered unrelated. The same dataset was used in earlier work [Baldi et al. Vullo & Frasconi. The DIPRO2 contains 567 such sequences. and the number of bonds is between 2 and 5 (since the case of 1 bond is trivial). 2004. 2001. 2004. We split SP39 into 4 different subsets. 2004]. 2000].html . 2004. 2000]. After redundance reduction there are a total of 1018 sequences. we adopt the same dataset splitting procedure as the one used in previous work [Fariselli & Casadio. release 39.. so we are able to perform learning on over 90% of all proteins. although the procedure had less effect due to the redundance reduction applied by the authors of the 2 http://contact. SP39 contains 446 sequences containing between 2 and 5 bonds. The SP39 dataset is extracted from the Swiss-Prot database of proteins [Bairoch & Apweiler. gap penalty 12 and gap extension 4). Even though our method is applicable to both sequences with a high number of bonds or sequences in which the bonding state of cysteine residues is unknown.edu/bridge.6 Experiments We assess the performance of our method on two datasets containing sequences with experimentally verified bonding patterns: DIPRO2 and SP39.

The features we experimented with were all based on the local regions around candidate cysteine pairs. is the same as SEQUENCE. xlm ) = exp( xjk −xlm γ 2 ). The second model. PROFILE. 10. 1997]. in which the entries denote whether or not a particular amino acid occurs at the particular position.10. The final set of features for each {j. we use multiple sequence alignment profile information. As in Baldi et al. For each pair of candidate cysteines {j. it is more efficient due to the low constants in the exponential problem size. We use the exponential sized represen- tation of Sec..5 and an RBF kernel K(xjk . except that instead of using the actual protein sequence. Thus. . We are currently working on an implementation of the certificate formulation Sec. 10.4 to handle longer sequences and non-perfect matchings (when bonding state is unknown). augmented with the linear distance between the cysteine residues. the 21st entry in the vector represents whether or not the amino-acid Alanine occurs at position 2 of the local window. where j < k. using a sequential optimization procedure which solves QP subproblems associated with blocks of training examples. the input at each position of a local window is the frequency of occurrence of each of the 20 amino-acids in the alignments. the actual sequence is expanded to a 20 × n binary vector. The first model. we extract the amino-acid sequence in windows of size n centered at j and k. k}. 10]. [2004]. SEQUENCE. k} pair of cysteines is simply the two local windows concatenated together. Multiple alignments were computed by running PSI-BLAST using default settings to align the sequence with all sequences in the NR database [Altschul et al. EXPERIMENTS 155 dataset. The models below use windows of size n = 9. with γ ∈ [0. Training time took around 70 minutes for 450 examples. we augment the features of each model with the number of residues between j and k. counting from the left end of the window.6. We used commercial QP software (CPLEX) to train our models. 10. For example.1. Models The experimental results we report use the dual formulation of Sec. uses the features described above: for each window.3 since for the case of proteins containing between two and five bonds. we describe several models we used. Below.

52 0. In each row.75 0.21 0.74 0. we show the performance gained by using multiple alignment information by comparing the results of the SEQUENCE model with the PROFILE. Results and discussion We evaluate our algorithm using two metrics: accuracy and precision. As we can see from Table 10.44 / 0.44 / 0. the best performance is in bold.51 0.27 Table 10.11 K 2 3 4 5 SEQUENCE 0. [2004].43 / 0.1: Numbers indicate Precision / Accuracy.156 CHAPTER 10. whereas precision measures the number of correctly predicted bonds as a fraction of the total number of possible bonds.79 / 0. preliminary results of the DAG-RNN model [Baldi et al..70 0.59 0. 2004] which represent the best currently published results. The DSSP program produces 8 types of secondary structure.67 / 0. (a) Performance of PROFILE model on SP39 vs. . as well as a length n binary vector representing the solvent accessibility at each position. PROFILE-SS.06 PROFILE 0. so we augment each local window of size n with an additional length 8 × n binary vector.17 (b) PROFILE-SS 0. PROFILE-SS models on the DIPRO2 dataset. As Table 10.70 / 0.61 / 0.73 / 0.60 / 0. We perform 4-fold crossvalidation on SP39 in order to replicate their setup.48 0. The first set of experiments compares our model to preliminary results reported in Baldi et al. The accuracy measure counts how many full connectivity patterns were predicted correctly.79 0.43 / 0. with similar or better levels of precision for all bond numbers. and slightly lower accuracies for the case of 2 and 3 bonds. In another experiment.62 / 0.46 / 0.73 0.16 (a) DAG-RNN 0.1(b).74 / 0. The third model.70 / 0. which represent the current top-performing system.69 0. (b) Performance of SEQUENCE.1 shows. PROFILE.24 0.29 / 0. the evolutionary information captured by the amino-acid alignment frequencies plays an important role in increasing the performance of the algorithm.62 / 0. the PROFILE model achieves comparable results.75 / 0.44 0.56 0.41 / 0.27 0. MATCHINGS K 2 3 4 5 PROFILE 0. augments the PROFILE model with secondary structure and solvent-accessibility information.74 / 0.59 / 0.

3: Given a set of nominal weights w0 and a perfect matching M . we examine the role that secondary structure and solvent-accessibility information plays in the model PROFILE-SS.1(b) shows that the gains are significant. which is not a maximum one with respect to w0 .3: Performance of PROFILE model as training set size changes for proteins with (a) 2 and 3 bonds (b) 4 and 5 bonds. Table 10. perhaps through more complex kernels. Such behavior is very promising.7. RELATED WORK 157 2 and 3 bonds 100% 90% 80% 70% 60% 50% 40% 30% 20% 100 150 200 250 300 Training Set Size 350 Accuracy (2 Bonds) Precision (3 Bonds) Accuracy (3 Bonds) 4 and 5 bonds 85% 75% 65% 55% 45% 35% 25% 15% 5% Precision (4 Bonds) Accuracy (4 Bonds) Precision (5 Bonds) Accuracy (5 Bonds) 400 450 100 150 200 250 300 Training Set Size 350 400 450 (a) (b) Figure 10. They do not provide a compact optimization problem for this related but . 4.3 shows the performance of the PROFILE model as training set size grows. The same phenomenon is observed by Vullo and Frasconi [2004] in their comparison of sequence and profile-based models. since it validates the applicability of our algorithm as the availability of high-quality disulfide bridge annotations increases with time.10. 10. find a new weight vector w that makes M optimal and minimizes ||w0 − w||p for p = 1. especially for sequences with 3 and 4 bonds. 10. As a final experiment. ∞.7 Related work The problem of inverse perfect matching has been studied by Liu and Zhang [2003] in the inverse combinatorial optimization framework we describe in Sec. We can see that for sequences of all bond numbers. Fig. both accuracy and precision increase as the amount of data grows. The trend is more pronounced for sequences with 4 and 5 bonds because they are sparsely distributed in the dataset. This highlights the importance of developing even richer features.

Residue-to-residue contacts (which include disulphide bridges) are predicted by solving a series of constrained integer programming problems.. [2002]. The prediction problem is solved using a weighted graph matching based on these probabilities. and uses a recursive neural network architecture [Frasconi et al. relying instead on the ellipsoid method with constraint generation. It also uses multiple alignment profile information and demonstrates its benefits over sequence information. It also improves on previous methods by not assuming knowledge of bonding state. Interestingly. [2004]. and edge weights correspond to attraction strength. The problem of learning the edge weights was addressed using a simulated annealing procedure. The method in Vullo and Frasconi [2004] takes a different approach to the problem. A different approach to predicting disulfide bridges is reported in Klepeis and Floudas [2003]. At prediction time the pattern scores are used to perform an exhaustive search on the space of all matchings. Their method uses Directed Acyclic Graph Recursive Neural Networks [Baldi & Pollastri. and HMMs . especially for the case of 2 and 3 disulfide bonds.158 CHAPTER 10. MATCHINGS different task. where bond prediction occurs as part of predicting β-sheet topology in proteins. the approach can be used to predict disulfide bonds with no knowledge of bonding state. Their method is only applicable to the case when bonding state is known. They modeled the prediction problem as finding a perfect matching in a weighted graph where vertices represent bonded cysteine residues. Their method performs better than the one in Vullo and Frasconi [2004] and is also the only one which can cope with sequences with more than 5 bonds. SVMs. In Baldi et al. but the results are not comparable with those in other publications. the authors switch to using a neural network for learning edge weights and achieve better performance. 2003] to predict bonding probabilities between cysteine pairs. It scores candidate connectivity patterns according to their similarity with respect to the correct pattern. 1998] to score candidate patterns. The method is computationally limited to sequences of 2 to 5 bonds. the authors achieve the current state-of-the-art performance on the task. In Fariselli et al. The problem of disulfide bond prediction first received comprehensive computational treatment in Fariselli and Casadio [2001]. The task of predicting whether or not a cysteine is bonded has also been addressed using a variety of machine learning techniques including neural networks.

We present two formulations: one which is based on a linear programming approach to matching. 2000. 10. will affect performance.. 2002.8. 2003] Currently the top performing systems have accuracies around 85%.8 Conclusion In this chapter. Our approach learns a parameterized scoring function that reproduces the observed matchings in a training set. requiring an exponential number of constraints. . While we have developed the framework to handle that task. 2004]. or extended local neighborhoods of the protein sequence. especially in comparison to existing methods [Baldi et al. and one which develops a certificate of matching optimality for a compact polynomial-sized representation. We also hope to explore the more challenging problem of disulfide connectivity prediction when the bonding state of cysteines is unknown. such as convolution kernels for protein sequences. it remains to experimentally determine how well the method performs. Frasconi et al.. 1999. we derive a compact convex quadratic program for the problem of learning to match. formulated as a weighted matching problem. Fiser & Simon. We apply our framework to the task of disulfide connectivity prediction. it will be worthwhile to examine how other kernels.. which have already addressed the more challenging setting. 2002.. Furthermore. the use of kernels makes it easy to incorporate rich sets of features such as secondary structure information.10. Our experimental results show that the method can achieve performance comparable to current top-performing systems.. Ceroni et al. In the future. Martelli et al. CONCLUSION 159 [Fariselli et al.

Emanuel & Fiat. in machine learning. encouraging those vertices to belong to a common cluster.. one user may organize her email messages by project and time. negative weights encourage the vertices to belong to different clusters. both of these parameters are effectively chosen to be the best possible by the problem definition. These properties make correlation clustering a promising approach to many clustering problems.Chapter 11 Correlation clustering Data can often be grouped in many different reasonable clusterings. another by sender and topic. It is formulated as a vertex partitioning problem: Given a graph with real-valued edge scores (both positive and negative). Demaine & Immorlica. Learning to cluster considers the problem of finding desirable clusterings on new data. it has been successfully applied to coreference resolution for proper nouns [McCallum & Wellner. 2003. For a given application. We focus on correlation clustering. For example. there might be only one of these clusterings that is desirable. a novel clustering method that has recently enjoyed significant attention from the theoretical computer science community [Bansal et al. and minimize the weight of inter-cluster edges. 2002. correlation clustering does not require the user to specify the number of clusters nor a distance threshold for clustering. Unlike most clustering formulations. partition the vertices into clusters to maximize the score of intra-cluster edges. 2003]. Positive edge weights represent correlations between vertices. given example desirable clusterings on training data. Images can be segmented by hue or object boundaries. several algorithms based on linear programming and positive-semidefinite 160 . 2003]. Recently.

1. but this will not generally produce a valid partition. We assume that the graph is fully connected (if it is not. we minimize the disagreement: the weight of positive excluded edges minus the weight of negative included edges. that represent whether node j and k belong to the same cluster. which will be constrained to have the same value. 11. Bansal et al. we can make it fully connected by adding appropriate edges jk with sjk = 0). Experiments on synthetic and real-world data show the ability of the algorithm to learn an appropriate clustering metric for a variety of desired clusterings. CLUSTERING FORMULATION 161 programming relaxations have been proposed to approximately solve this problem. sjk yjk . one for each edge jk. we maximize the “agreement” of the partition with the positive/negative designations: the weight of the positive included edges minus the weight of negative excluded edges.11. E) with N nodes and edge score sjk for each jk in E. jk:sjk <0 (M AX AGREE) (M IN D IS AGREE) min y∈Y jk:sjk >0 sjk (1 − yjk ) − The motivation for the names of the two problems comes from separating the set of edges into positive weight edges and negative weight edges.1 Clustering formulation An instance of correlation clustering is specified by an undirected graph G = (V. In . we employ these relaxations to derive a max-margin formulation for learning the edge scores for correlation clustering from clustered training data. and fix them to have value 1 and set sjj = 0. (j < k). We formulate the approximate learning problem as a compact convex program with quadratic objective and linear or positive-semidefinite constraints. [2002] consider two related problems: max y∈Y jk:sjk >0 sjk yjk − jk:sjk <0 sjk (1 − yjk ). we introduce both yjk and ykj variables. For notational convenience. In M IN D IS AGREE. We define binary variables yjk . The best score is obviously achieved by including all the positive and excluding all the negative edges. Let Y be the space of assignments y that define legal partitions. In M AX AGREE. We also introduce yjj variables. In this chapter.

k = . With these definitions. s. l. The optimal partition for the two problems is of course the same (if it is unique). k. yjj = 1. ∀j. Bansal et al. ∀j. l = b and j = b. (11. ∀j. s− = jk:sjk <0 sjk . s+ = jk:sjk >0 sjk . M AX AGREE.2) yjk ∈ {0. To see this. yjl + ykl ≤ yjk + 1.t.1 Linear programming relaxation In order to insure that y defines a partition. 1}. transitive binary relation induces a partition of the objects. 2003]. it is sufficient to enforce a kind of triangle inequality for each triple of nodes j < k < l: yjk + ykl ≤ yjl + 1. Demaine & Immorlica. 2003.1) The triangle inequality enforces transitivity: if j and k are in the same cluster (yjk = 1) and k and l are in the same cluster (ykl = 1). 11. Note that the constraints imply that yab = yba for any two nodes a and b. k = b. which we consider in the next sections. we can express the M AX AGREE problem as an integer linear program (ignoring the constant −s− ): max jk sjk yjk yjk + ykl ≤ ylj + 1. The other two cases are similar. k. Then the value of M AX AGREE is s∗ − s− and the value of M IN D IS AGREE s+ − s∗ . CORRELATION CLUSTERING particular.162 CHAPTER 11.. consider the inequalities involving node a and b with j = a. Several approximation algorithms have been developed based on Linear and Semidefinite Programming [Charikar et al. 2003. We will concentrate on the maximization version.1. let s∗ = max y∈Y sjk yjk . then j and l will be forced to be in this cluster (yjl = 1). yjk + yjl ≤ ykl + 1. (11. [2002] show that both of these problems are NP-hard (but have different approximation hardness). Any symmetric. Emanuel & Fiat.

2 Semidefinite programming relaxation An alternative formulation [Charikar et al. vK }.1. Choose a collection of orthogonal unit vectors {v1 . µjk ≥ 0. l = a: yab + ybb ≤ yba + 1. ∀j.4) s. l. consider any clustering solution. . [2003] show that the integrality gap of this upper bound is at least 2/3: s∗ − s− 2 < . ∀j. To motivate this relaxation.t. . one for each cluster in the solution.3) Note that µjk ≤ 1 is implied since the triangle inequality with j = l gives µjk + µkj ≤ µjj + 1. µjj = 1. The score of the clustering . Since yaa = ybb = 1. µjk ≥ 0. we have µjk ≤ 1. k. Let mat(µ) denote the variables µjk arranged into a matrix.11. CLUSTERING FORMULATION 163 a. k. max jk sjk µjk mat(µ) 0.2) with continuous variables 0 ≤ µjk ≤ 1. The LP relaxation is obtained by replacing the binary variables yjk ∈ {0. ∀j. The LP solution. yba + yaa ≤ yab + 1. Charikar et al. Every vertex j in the cluster is assigned the unit vector vj corresponding to the cluster it is in. In effect. µjk + µkl ≤ µlj + 1. and since µjj = 1 and µjk = µkj . ∀j. k. (11. µjj = 1. 2003] is the SDP relaxation. sLP . vj vk = 0. we substituted the triangle inequalities by the semidefinite constraint. if not. (11. . We are not guaranteed that this relaxation will produce integral solutions. (11.1. then vj vk = 1. LP − s− s 3 11. 1} in Eq.. is an upper bound on the s∗ . If vertices j and k are in the same cluster. ∀j. we have yab = yba . .

(11. features might be distances along different dimensions. Furthermore.164 CHAPTER 11. we have mat(µ) mat(µ) = j 0.jk (1) + (1 − yjk ) i.jk (yjk ): = jk i. we assume that the loss function decomposes over the edges.jk . where i.jk (0) = i (0) + jk yjk i.2 Learning formulation The score for an edge is commonly derived from the characteristics of the pair of nodes.82843: s∗ − s− ≤ 0. sSDP . we parameterize the score as a weighted combination of basis functions sjk = w f (xjk ). The vectors generating the inner products are unit vectors by the requirement µjj = 1.jk (0). Hence we write w f (x.82843.jk (1) − i. where xjk is a set of features associated with nodes j and k. w. while if one is clustering points in IRn . In document clustering. For example. the entries of fxjk might be the words shared by the documents j and k.jk = i.jk (yjk ) = jk yjk i. is also an upper bound on the s∗ and Charikar et al. f (xjk ) ∈ IRn . y) = jk yjk w f (xjk ). which can be decomposed into a sum of outer products vj vj . the Hamming loss counts the number of . they do not necessarily form a set of orthogonal vectors. However. The SDP solution.4). CORRELATION CLUSTERING solution can now be expressed in terms of the dot products vj vk . Specifically. The entries µjk correspond to inner products vj vk . [2003] show that the integrality gap of this upper bound is at least 0. into a sum of edge losses i (y) i. We assume f (xjj ) = 0 so that sjj = 0. In the SDP relaxation in Eq. sSDP − s− 11.

(11.5) We can now use the LP relaxation in Eq. s.t.5). l.jkl for each triangle inequality and zi. We use these upper-bounds in the min-max formulation to achieve approximate max-margin estimation. (11.ljk − λi.j for the identity on the diagonal.jkl + j zi.jkl + λi. k.2. λi. (11.jk . (11.ljj − λi.jlj ] ≥ i.jj . (11.3) and the SDP relaxation in Eq.j (i) i.klj ] ≥ w f (xjk ) + zi. we introduced a dual variable λi. l [λi. LEARNING FORMULATION 165 edges incorrectly cut or uncut by a partition.jk = 1 − 2yjk .6) ∀j = k. The dual of the LP based upper bound for example i is i (0)+ min jkl λi. ∀j. the loss augmented maximization is i (0) + max y∈Y jk yjk [w f (xjk ) + (i) i.jk ].jk . ∀j.jkl ≥ 0. H i (y) = jk 1 jk = yjk ) = I(y jk (i) (i) yjk + jk (i) yjk (1 − 2yjk ) = (i) H i (0) + jk yjk i. Above.j + l [λi.4) as upper bounds on Eq.jjl + λi. .11. With this assumption. Note the righthand-side of the second set of inequalities follows from the assumption f (xjj ) = 0. where i.

[λi. λi. ∀j.jj ≥ mat(λi ) i. we have: min s.jk ≥ w f (xjk ) + zi.9) zi.jj (i) i. ∀j = k.jj .j − λi. k. the dual of the SDP based upper bound is i (0)+ min j zi.ljj − λi. y(i) ) + ξi ≥ i (0) + j (11.j −λi.j . 1 ||w||2 + C ξi 2 w f (x(i) . i. ∀i.t. −λi.jkl + j (i) f (xjk ) i.8) ∀j = k. ∀i. Plugging the SDP dual into the min-max formulation. s.jk . (i) i.ljk − λi. l.jk . zi. 0. k. ∀j. ∀j. i. .jj . (11. l. ∀i. y(i) ) + ξi ≥ i (0) + jkl (11. ∀i.jkl ≥ 0.t.jk ≥ w f (xjk ) + zi. ∀j. zi. ∀i.j .jlj ] ≥ ∀i.7) λi.jj ≥ mat(λi ) 0.j + l [λi. ∀j.j − λi.t.jk . we have: min s.166 CHAPTER 11. Similarly. 1 ||w||2 + C ξi 2 w f (x(i) . ∀i.jkl + λi.jjl + λi. ∀j = k. CORRELATION CLUSTERING Plugging this dual into the min-max formulation.klj ] ≥ w l + ∀i.

(11.jk ≥ 0.jk 1 i. l are replaced by the corresponding mat(µi ) µi. mat(µi ) µi.7): max i. for example. max i.jj = 1. ∀i.jk − C 2 2 (i) (yjk i. we can use kernels K(xjk . (11. Here we give the dual of Eq. k.jk − (i) µi. m). (j. k. Therefore. The dual objective can be expressed in terms of dot-products f (xjk ) f (xlm ).jk µi. ∀j. (11. ∀i. ∀i. ∀j. If we are clustering points in Euclidian space. µi.jk ≥ 0. ∀i.jj = 1.9) is very similar. ∀j. ∀j. 2 (i) (yjk i.jk − C 2 0. (i) (i) (11. the kernel could be a function of the two segments corresponding to the pairs of points. (11. ∀i. The dual of Eq.7) and Eq. k.jk 0: − (i) µi.jk )f (xjk ) s.3. k) and (l. The relation between the primal and dual solution is w = C i. k. ∀j. µi.jk 1 i.t.11. µjk + µkl ≤ µlj + 1.jk (yjk − µjk )f (xjk ). ∀i. l.9) provide some insight into the structure of the problem and enable efficient use of kernels. a polynomial kernel over their lengths and angle between them. and measures the similarity between the relation between the nodes of each pair.jk )f (xjk ) s. DUAL FORMULATION AND KERNELS 167 11. .3 Dual formulation and kernels The dual of Eq. µi.t. ∀j. This kernel looks at two pairs of nodes. xlm ) to define the space of basis functions.jk ∀i.10) One important consequence of this relationship is that the edge parameters are all support vector expansions. except that the linear transitivity constraints µjk + µkl ≤ µlj + 1.

This first dimension is thus the relevant feature. we explore on a synthetic example how our algorithm deals with irrelevant features. we generate data (100 points) from a mixture of two onedimensional Gaussians. 2 11.4. 11. We used a basis function for each dimension fd (xjk ) = e−(xj [d]−xk [d]) .4. Let xj denote each point and xj [d] denote the d-th dimension of the point.168 CHAPTER 11. which the users had manually filed into different folders. The noise is independently generated for each dimension. 1 http://www. 11. email clustering and image segmentation. from a mixture of Gaussians with same difference in means and variance as for the relevant dimension. CORRELATION CLUSTERING 11. with an average of 3-4 folders. We gathered the data from the SRI CALO project.1 Our dataset consisted of email from seven users (approximately 100 consecutive messages per user). and two real data sets. Figure 11. We are interested in the problem of learning to cluster emails.1 Irrelevant features In this section. In particular. plus an additional constant basis function. Random clustering will give an accuracy of 50%.sri. The number of folders for each users varied from two to six. The results in Fig. The comparison with k-means is simply a baseline to illustrate the effect of the noise on the data. Then we add noise components in D additional (irrelevant) dimensions. The accuracy is the fraction of edges correctly predicted to be between/within cluster.4 Experiments We present experiments on a synthetic problem exploring the effect of irrelevant basis functions (features).1(b) illustrate the capability of our algorithm to learn to ignore irrelevant dimensions.com/project/CALO .1(a) shows the projection of a data sample onto the first two dimensions and on two irrelevant dimensions.2 Email clustering We also test our approach on the task of clustering email into folders. The training and test data consists of a 100 samples from the model. where each mixture component corresponds to a cluster.ai.

11. averages are over 20 runs. The algorithm is therefore able to automatically learn the relative importance of certain email fields to filing two messages together.4. Error-bars denote one standard deviation.75 0. Cc. Also. to. We compare our method to the k-means clustering algorithm using with the same word .85 0. what score function sjk causes correlation clustering to give clusters similar to those that human users had chosen? To test our learning algorithm. irrelevant ones.95 test accuracy 0.65 0. One feature was used for each common word in the pair of emails (except words that appeared in more than half the messages. (c) Performance on 2 cluster problem as function of the number of irrelevant noise dimensions. The basis functions f (xjk ) measured the similarity between the text of the messages. including the from. Learning to cluster is the solid line. as well as importance of meaningful words versus common. Specifically. “To:” field. we use each user as a training set in turn. additional features captured the proportion of shared tokens for each email field.45 1 2 4 8 16 32 64 128 lcc k-means 2 1 0 −1 −2 −3 −4 −4 4 −3 −2 −1 0 1 2 3 3 2 1 0 −1 −2 # noise dimensions −3 −2 −1 0 1 2 3 4 −3 −4 −4 (a) (b) Figure 11. which were deemed “stop words” and omitted). subject and body fields. Accuracy is the fraction of edges correctly predicted to be between/within cluster. the similarity between the “From:” field. EXPERIMENTS 169 4 3 0. k-means the dashed line.1: (a) Projection onto first two dimensions (top) and two noise dimensions (bottom). We then use the learned parameters to cluster the other users’ mail. learning the parameters from the partition of a single user’s mailbox into folders.55 0. and “Cc:” field.

Fig. We also informed our algorithm of the number of clusters by uniformly adding a positive weight to the edge weights to cause it to give the correct number of clusters. features.3 0.2: Average Pair F1 measure for clustering user mailboxes.4.8 0.6 0. For example. 11. based on the hue and intensity. and took the best clustering out of five tries.3(a) and (b) show two distinct segmentations: one very coarse. we may prefer the first over the second or vice-versa. Depending on the task at hand.170 CHAPTER 11.2 show the average F 1 measure (harmonic mean of precision and recall. 2001] for which two users had significantly different segmentations. Our algorithm significantly outperforms k-means on several users and does worse only for one of the users. It is precisely this kind of variability in the similarity judgements that . We made this comparison somewhat easy for k-means by giving it the correct number of clusters k. CORRELATION CLUSTERING 0.5 Pair F1 lcc k-means 0.2 0.4 0. and one much finer. mostly based of overall hue. 11.. 1999]) computed on the pairs of messages that belonged to the same cluster. We selected images from the Berkeley Image Segmentation Dataset [Martin et al.7 0. a standard metric in information retrieval [Baeza-Yates & Ribeiro-Neto.1 0 1 2 3 4 User 5 6 7 Figure 11.3 Image segmentation We also test our approach on the task of image segmentation. We performed a simple binary search on this additional bias weight parameter to find the number of clusters comparable to k. 11. The results in Fig.

3(a) and the other using Fig. In order to segment the image. we used a fairly simple set of features that are very easy to compute. We then selected about a hundred largest regions. We trained two models.B channel). 11. These regions are the objects that our algorithm learns to cluster. we can consider their shape.3(b) and . etc. distance. (We then use the learned metric to greedily assign the remaining small regions to the large adjoining regions. color distribution. which covered 80-90% of the pixels. we want our algorithm to capture.11. one using Fig. we calculated the bounding box. We then computed three distances (one for each HSV channel). area and average color (averaging the pixels in the RGB space). presence of edges between them.4. EXPERIMENTS 171 (a) (b) Figure 11. In our experiments. We connected two adjacent pixels by an edge if their RGB value was the same at a coarse resolution (4 bits per each of the R. All features were normalized to have zero mean and variance 1. we first divided it contiguous regions of approximately the same color by running connected components on the pixels.3: Two segmentations by different users: training image with (a) coarse segmentation and (b) fine segmentation.G. 11.) There is a rich space of possible features we can use in our models: for each pair of regions. For each region. as well as the distance in pixels between the bounding boxes and the area of the smaller of the two regions.

we will instead seek to learn a good metric for the clustering algorithm. 2003]. An example of work in a similar vein is the algorithm for learning a distance metric for spectral clustering [Bach & Jordan. 2003]. Thus.4(b) and (c).4(a). 11. In contrast to algorithms that learn a metric independently of the algorithm that will be used to cluster the data. Bar-Hillel et al. Note that mountains. 2002.. The results are shown in Fig. tested on the image in Fig. 11.172 CHAPTER 11. which is more efficient than correlation clustering.5 Related work The performance of most clustering algorithms depends critically on the distance metric that they are given for measuring the similarity or dissimilarity between different datapoints.4: Test image: (a) input. CORRELATION CLUSTERING (a) (b) (c) Figure 11. (b) segmentation based on coarse training data (c) segmentation based on fine training data. Recently. However the objective in the learning . 11. for which we use LP or SDP formulations. respectively. The clustering algorithm essentially uses an eigenvector decomposition of an appropriate matrix derived from the pairwise affinity matrix. instead of using an externally-defined criterion for choosing the metric. tuning one to the other in a joint optimization. rocks and grass are segmented very coarsely based on hue in (b) while the segmentation in (c) is more detailed and sensitive to saturation and value of the colors. a number of algorithms have been proposed for automatically learning distance metrics as a preprocessing step for clustering [Xing et al.. we describe a formulation that tightly integrates metric learning with the clustering algorithm.

On the theoretical side.6 Conclusion We looked at correlation clustering. it would be interesting to work out a PAC-like bound for generalization of the learned score metric. 11. Our approach ties together the inference and learning algorithm. We showed results on a synthetic dataset. The main limitation of the correlation clustering is scalability: the number of constraints (|V|3 ) in the LP relaxation and the size of the positive-semidefinite constraint in the SDP relaxation. .11. It would be very interesting to explore constraint generation or similar approaches to speed up learning and inference. showcasing robustness to noise dimensions. and how to learn the edge weights from example clusterings.6.. Experiments on the CALO e-mail and image segmentation experiments show the potential of the algorithm on realworld data. and attempts to learn a good metric specifically for the clustering algorithm. CONCLUSION 173 formulation proposed in Bach and Jordan [2003] is not convex and difficult to optimize.

174 CHAPTER 11. CORRELATION CLUSTERING .

Part IV Conclusions and future directions 175 .

For some of these models. 3D terrain classification. parameterized scoring function w f (x. y : g(x. Our class of models has the following linear form: hw (x) = arg max w f (x. alternative estimation methods are intractable. a model defines a compact.y)≤0 176 . Fundamentally. including handwriting recognition. y). Essentially. hypertext categorization. 12.Chapter 12 Conclusions and future directions This thesis presents a novel statistical estimation framework for structured models based on the large margin principle underlying support vector machines. we rely on the expressive power of convex optimization problems to compactly capture inference or solution optimality in structured models. Directly embedding this structure within the learning formulation produces compact convex problems for efficient estimation of very complex models. The framework results in several efficient learning formulations for complex prediction tasks. We develop theoretical foundations for our approach and show a wide range of experimental applications. natural language parsing.1 Summary of contributions We view a structured prediction model as a mapping from the space of inputs x ∈ X to a discrete vector output y ∈ Y. disulfide connectivity in protein structure prediction. email organization and image segmentation. y) and prediction using the model reduces to finding the highest scoring output y given the input x.

which is generally exponential in the number of variables in each y(i) . SUMMARY OF CONTRIBUTIONS 177 where w ∈ IRn is the vector of parameters of the model. y) ∈ IRn represent salient features of the input/output pair. ∀y ∈ Y (i) . y∈Y (i) where Y (i) = {y : g(x(i) . min s. 10). This definition covers a broad range of models. from probabilistic models such as Markov networks and context free grammars to more unconventional models like weighted graph-cuts and matchings.y)≤0 w f (x. ∀i. y(i) )}m . y) ∈ IRk define the space of feasible outputs y given the input x and basis functions f (x. y) can be solved (or closely approximated) by an efficient algorithm that exploits the structure of the constraints g and basis functions f . constraints g(x. we assume that the inference problem arg maxy : g(x. 1 ||w||2 2 w fi (y(i) ) ≥ w fi (y) + i (y).12. . we develop methods for finding parameters w such i=1 that: arg max w f (x(i) . The naive formulation1 uses i |Y (i) | linear constraints. 7 and Ch. we omit the slack variables in this summary. y) ≤ 0}. ∀i.1 Structured maximum margin estimation Given a sample S = {(x(i) . y) ≈ y(i) . 12. Although the space of outputs {y : g(x. y) ≤ 0} is usually immense. In many models where maximum likelihood estimation is intractable. we provide exact maximum margin solutions (Ch.1.1. We propose two general approaches that transform the above exponential size QP to an exactly equivalent polynomial size QP in many important classes of models. These formulations allow us to find globally optimal parameters (with fixed precision) in polynomial time using standard optimization software.t. 1 For simplicity.

bipartite matchings. we can embed the maximization inside the min-max formulation to get a compact convex program equivalent to the naive exponential formulation.t.. For example.g. can be solved in polynomial time using combinatorial algorithms. 2003]. ∀i. In some cases. though. and many other models. SDP. 1997. etc. QP. y . y∈Y (i) The key to solving the estimation problem above efficiently is the loss-augmented inference problem maxy∈Y (i) [w fi (y) + i (y)]. We show that this approach leads to exact polynomial-size formulations for estimation of low-treewidth Markov networks. Even if maxy∈Y (i) w fi (y) can be solved in polynomial time using convex optimization. we can find a compact certificate of optimality that guarantees that y(i) = arg max[w fi (y) + i (y)]. however. associative Markov networks over binary variables. maximum weight perfect (non-bipartite) matching and spanning tree problems can be expressed as linear programs with exponentially many constraints. Both of these problems. We show that if we can express the (loss-augmented) inference as a compact convex optimization problem (e.178 CHAPTER 12. Certificate formulation There are several important combinatorial problems which allow polynomial time solution yet do not have a compact convex optimization formulation.). the form of the loss term i (y) is crucial for the loss-augmented inference to remain tractable. 1 ||w||2 2 w fi (y(i) ) ≥ max [w fi (y) + i (y)]. context-free grammars. We typically use a natural loss function which is essentially the Hamming distance between y(i) and h(x(i) ): the number of variables predicted incorrectly. CONCLUSIONS AND FUTURE DIRECTIONS Min-max formulation We can turn the above problem into an equivalent min-max formulation with i non-linear max-constraints: min s. Schrijver. LP. but no polynomial formulation is known [Bertsimas & Tsitsiklis.

1. 10). Altun et al. and does not require a normalized distribution P (y | x) over all outputs. maximum margin estimation can be tractable when maximum likelihood is not. 1993. This observation allows us to apply our framework to an even broader range of models with combinatorial structure than the min-max formulation. maximum likelihood There are several theoretical advantages to our approach in addition to the empirical accuracy improvements we have shown experimentally. the solutions to the estimation are relatively sparse in the dual space (as in SVMs). many other problems in which inference is solvable by dynamic programming). Lafferty et al. we can derive a compact certificate for the spanning tree problem. 1979. SUMMARY OF CONTRIBUTIONS 179 For perfect (non-bipartite) matchings. our approach has an additional advantage. Zhu & Hastie. to learn a probabilistic model P (y | x) over bipartite matchings using maximum likelihood requires computing the normalizing partition function. which is #P-complete [Valiant. 2004]. 2001..12. Maximum margin vs. Garey & Johnson. By contrast. this certificate is a condition that ensures there are no augmenting alternating cycles (see Ch.. Maximum likelihood models with kernels are generally non-sparse and require pruning or greedy support vector selection methods [Wahba et al. 1979]. The certificate formulation relies on the fact that verifying optimality of a solution is often easier than actually finding one. This simple set of linear constraints scales linearly with the number of edges in the graph. Because of the hinge-loss. For example. In models that are tractable for both maximum likelihood and maximum margin (such as low-treewidth Markov networks. which makes the use of kernels much more efficient. context free grammars. Similarly. . Similar results hold for an important subclass of Markov networks and non-bipartite matchings. Because our approach only relies on using the maximum in the model for prediction. 2004. maximum margin estimation can be formulated as a compact QP with linear constraints.. We can express this condition by defining an auxiliary distance function on the nodes an a set of linear constraints that are satisfied if and only if there are no negative cycles.

◦ Low-treewidth Markov networks We use a compact LP for MAP inference in Markov networks with sequence and other low-treewidth structure to derive an exact. 7. 11.180 CHAPTER 12. 1977]. representational extensions. relational The largest portion of the thesis is devoted to novel estimation algorithms. Also. CONCLUSIONS AND FUTURE DIRECTIONS There are. Many such problems can be written as integer programs. QPs or SDPs often provide excellent approximation algorithms and fit well within our framework.. convex learning formulation. Approximations In many problems. it is very difficult to solve using standard software. Relaxations of such integer programs into LPs. The dual formulation allows efficient integration of kernels with graphical models that leverages rich high-dimensional representations for accurate prediction in real-world tasks. In applications where probabilistic confidence information is a must. We present an efficient algorithm for solving the estimation problem called Structured SMO. associative. particularly the min-max formulation. for large datasets (millions of examples). ◦ Scalable online algorithm Although our convex formulation is a QP with linear number of variables and constraints in the size of the data. for example. 12. multiway cuts in Ch. of course. compact. we consider MAP inference in large treewidth Markov networks in Ch.2 Markov networks: max-margin. 8. maximum likelihood is much more appropriate. generalization analysis and experimental validation for Markov networks. several advantages to maximum likelihood estimation. probabilistic interpretation permits the use of well-understood algorithms such as EM [Dempster et al.1. We show empirically that these approximations are very effective in many applications. in training settings with missing data and hidden variables. the maximization problem we are interested in may be NP-hard. Our online-style algorithm uses . graph-partitioning in Ch.

◦ Learning associative Markov networks (AMNs) We define an important subclass of Markov networks that captures positive correlations present in many domains. we provide an approximation that works well in practice. our framework allows exact estimation of networks of arbitrary connectivity and topology. ◦ Representation and learning of relational Markov networks We introduce relational Markov networks (RMNs). 1952] often used in computer vision and allows exact MAP inference in the case of binary variables. for which likelihood methods are believed to be intractable. for associative Markov networks over binary variables. relations. By constraining the class of Markov networks to AMNs. For the non-binary case. The graphical structure of an RMN is based on the relational structure of the domain. ◦ Generalization analysis We analyze the theoretical generalization properties of max-margin estimation in Markov networks and derive a novel margin-based bound for structured prediction.1. As a result.g. We show how to express the inference problem using an LP which is exact for binary networks. This is the first bound to address structured error (e. SUMMARY OF CONTRIBUTIONS 181 inference in the model and analytic updates to solve extremely large estimation problems. at run-time. our models are learned efficiently and. which compactly define templates for Markov networks for domains with relational structure: objects. . and can easily model complex interaction patterns over related entities. We apply this class of models to classification of hypertext using hyperlink structure to define relations between webpages. We use a compact approximate MAP LP in these complex Markov networks.12. to derive an approximate max-margin formulation. This class of networks extends the Potts model [Potts. We present an AMN-based method for object segmentation from 3D range data.. can scale up to tens of millions of nodes and edges by using graph-cut based inference [Kolmogorov & Zabih. proportion of mislabeled pixels in an image). in which exact inference is intractable. 2002]. attributes.

matching. graph partitions for clustering documents and segmenting images. The algorithm we propose uses kernels. tuning one to the other in a joint optimization. We show experimental evidence of the model’s improved performance over several baseline models. clustering The other large portion of the thesis addresses a range of prediction tasks with very diverse models: context free grammars for natural language parsing. perfect matchings for disulfide connectivity in protein structure prediction. to derive an exact. including email folder organization and image segmentation.1.3 Broader applications: parsing. ◦ Learning to cluster By expressing the correlation clustering problem as a compact LP and SDP. namely the absence of augmenting alternating cycles. we use the min-max formulation to learn a parameterized scoring function for clustering. CONCLUSIONS AND FUTURE DIRECTIONS 12. ◦ Learning to match We use a combinatorial optimality condition. We apply our framework to prediction of disulfide connectivity in proteins using perfect matchings. efficient certificate formulation for learning to match. In contrast to algorithms that learn a metric independently of the algorithm that will be used to cluster the data. ◦ Learning to parse We exploit dynamic programming decomposition of context free grammars to derive a compact max-margin formulation. Experiments on synthetic and real-world data show the ability of the algorithm to learn an appropriate clustering metric for a variety of desired clusterings. we describe a formulation that tightly integrates metric learning with the clustering algorithm. We formulate the approximate learning problem as a compact convex program. . which makes it possible to efficiently embed input features in very high-dimensional spaces and achieve state-of-the-art accuracy. We build on a recently proposed “unlexicalized” grammar that allows cubic time parsing and we show how to achieve high-accuracy parsing (still in cubic time) by exploiting novel kinds of lexical information.182 CHAPTER 12.

For example.d.1 Theoretical analysis and optimization algorithms ◦ Approximation bounds In several of the intractable models.. ◦ Problem-specific optimization methods Although our convex formulations are polynomial in the size of the data. In spatial prediction tasks.12. This distribution-free assumption leads to a worst case analysis. y). 7 and correlation clustering in Ch. while some assumptions about the approximate decomposition P (y | x) may be warranted. 12. like multi-class AMNs in Ch.e. 11.2 Extensions and open problems There are several immediate applications. An open question is to translate these error bounds on inference into error bounds on the resulting max-margin formulations. Similar assumptions may be made for some “degree of separation” in relational domains. More generally. and more general learning settings. 5. for sequential prediction problems.2.i. given the input and previous k labels. we used approximate convex programs within the min-max formulation. EXTENSIONS AND OPEN PROBLEMS 183 12. y(i) ) being i. less immediate extensions and open problems for our estimation framework. the next label is independent of the labels more than k in the past). scaling .2. the Markov assumption of some finite order is reasonable (i. These approximate inference programs have strong relative error bounds. including further theoretical analysis and new optimization algorithms. novel prediction tasks. without any assumption about the distribution of P (y | x). ◦ Generalization bounds with distributional assumptions In Ch. it would be interesting to exploit such conditional independence assumptions or asymptotic bounds on entropy of P (y | x) to get generalization guarantees even from a single structured example (x. we presented a bound on the structured error in Markov networks. a label variable is independent of the rest given a large enough ball of labels around it. relying only on the samples (x(i) . We organize these ideas into several sections below.

Our framework can be applied (just as in context free grammar estimation) to efficiently learn a more complex edit function that depends on the contextual string features. .2 Novel prediction tasks ◦ Bipartite matchings Maximum weight bipartite matchings are used in a variety of problems to predict mappings between sets of items. points on two shapes are matched based on their local contour features [Belongie et al. 1999. 1998]. we have presented the Structured SMO algorithm.2. Both algorithms rely on dynamic programming decompositions. 2000]. matchings. ◦ Continuous prediction problems We have addressed estimation of models with discrete output spaces. Similarly. such as graph-cuts. CONCLUSIONS AND FUTURE DIRECTIONS up to larger datasets will require problem-specific optimization methods. In machine translation. a string edit distance is used to determine which portions of the sequences align to each other [Needleman & Wunsch.184 CHAPTER 12. structured classification. matchings are used to map words of the two languages in aligned sentences [Matusov et al. create a need for new algorithms that can effectively use combinatorial optimization as a subroutine to eliminate the dependence on general-purpose convex solvers. 2004]. Leslie et al.. 1970. However. ◦ Sequence alignment In standard pairwise alignment of biological sequences. Lodhi et al... 2004]. Our framework provides an exact. perhaps using novel string kernels [Haussler. models which do not permit such decompositions. Another algorithm useful for such models is Exponentiated Gradient [Bartlett et al. efficient alternative to the maximum likelihood estimation for learning the matching scoring function. In 2D shape matching... generalizing classification models to multivariate. Finding the best alignment involves a dynamic program that generalizes the longest common subsequence algorithm. 2002. and many others. For lowtreewidth Markov networks and context free grammars. Durbin et al. we can consider a whole range of problems where the prediction variables are continuous. 12. 2002]..

convex programs are often used to model portfolio management. one could learn a user’s return projection and risk assessment function from observed portfolio allocations by the user. interconstrained real-valued outputs.3 More general learning settings ◦ Structure learning We have focused on the problem of learning parameters of the model (even though our kernelized models can be considered non-parametric). Details of this are beyond the scope of this thesis. involving correlated. Hamming distance equivalent. there are obstacles to directly applying the min-max or certificate formulations.2. not concave (it turns out that it is actually possible to use L∞ loss). The standard method of greedy stepwise selection followed by re-estimation is very . with stoichiometric constraints [Varma & Palsson. but it suffices to say that loss-augmented inference using. EXTENSIONS AND OPEN PROBLEMS 185 Such problems are a natural generalizations of regression. especially in spatial and relational domains. L1 loss (or L2 loss). there is a wealth of possible structures (cliques in the network) one can use to model a problem. 1991. 1997]. These problems are similar to the discrete structured prediction models we have considered: inference in the model can formulated as a convex optimization problem. For example. no longer produces a maximization of a concave objective with convex constraints since L1 . Developing effective loss functions and max-margin formulations for the continuous setting could provide a novel set of effective models for structured multivariate real-valued prediction problems. for example. It would be interesting to use observed equilibria data under different conditions to learn what “objective” the cell is maximizing. 1994]. Markowitz portfolio optimization is formulated as a quadratic program which minimizes risk and maximizes expected return under budget constraints [Markowitz. 12. It is particularly interesting to explore the problem of inducing these cliques automatically from data. several recent models of metabolic flux in yeast use linear programming formulations involving quantities of various enzymes.12. In financial modeling. In the case of Markov networks. In this setting. However. Luenberger.2. L2 are convex.

in machine translation. Bach & Jordan.. Corduneanu & Jaakkola. . Especially in the case of combinatorial structures. 1977. in handwriting recognition. 2001. This setting has been studied mainly in the probabilistic. It is possible that AMNs. while much of the easily accessible data is not at all or suitably labeled. 2001]. generative models often using the EM algorithm [Dempster et al. 1997. Recent work on selecting input features in Markov networks (or CRFs) uses several approximations to learn efficiently with millions of candidate features [McCallum. extensions of our framework allow opportunities for problem-specific convex approximations to be exploited. clique selection is still relatively unexplored. Zhu et al. Discriminative methods have been explored far less. Cowell et al. 2003]. There are several more general settings we would like to extend our framework. by restricting the network to be tractable under any structure. The simplest setting is a mix of labeled and unlabeled examples. For example.. where a small supervised dataset is augmented by a large unsupervised one. we may not have each letter segmented out but instead just get a word or sentence as a label for the entire image. There has been much research in this setting for classification [Blum & Mitchell. These correspondence variables can be treated as hidden variables. Similarly.186 CHAPTER 12. 2000.. 2002. Nigam et al. 1998. word correspondences between pairs of sentences are usually not manually annotated (at least not on a large scale). Szummer & Jaakkola. 2003]. the principle of regularizing (discouraging) decision boundaries near densely clustered inputs could be applicable to our structured setting. A more complex and rich setting involves presence of hidden variables in each example. ◦ Semi-supervised learning Throughout the thesis we have assumed completely labeled data. Although most of this work has been done in a probabilistic setting.. Chapelle et al.. CONCLUSIONS AND FUTURE DIRECTIONS expensive in general networks [Della Pietra et al. may permit more efficient clique selection methods.. 1999]. However. 2003. This assumption often limits us to relatively small training sets where data has been carefully annotated.

We hope that continued research in this framework will help tackle evermore sophisticated prediction problems in the future.12. . Our approach has several theoretical and practical advantages over standard probabilistic models and estimation methods for structured prediction.3 Future We have presented a supervised learning framework for a large class of prediction models with rich and interesting structure. FUTURE 187 12.3.

The covering number is a key quantity in measuring function complexity. We reproduce the relevant definitions and theorems from Zhang [2002] here to highlight the necessary extensions for structured classification. 188 . Here. Intuitively. z) = w z where the norms of the vectors w and z are bounded.) Zhang provides generalization guarantees for linear binary classifiers of the form hw (x) = sgn(w x).1 uses the covering number bounds of Zhang [2002] (in the Data-Dependent Structural Risk Minimization framework [Shawe-Taylor et al. the covering number of an infinite class of functions (e. parameterized by a set of weights w) is the number of vectors necessary to approximate the values of any function in the class on a sample.g. His analysis is based on the upper bounds on the covering number for the class of linear functions FL (w.5.1 Proof of Theorem 5.Appendix A Proofs and derivations A.. Margin-based analysis of generalization error uses the margin achieved by a classifier on the training set to approximate the original function class of the classifier by a finite covering with precision that depends on the margin.5. 1998]. we will only define the ∞-norm covering number.1 The proof of Theorem 5.

so that 0-1 loss of I(v sgn(w x) is f (FL (w. v ) = |v − v |.v )<2γ f (v ) be the γ-margin loss for a metric ρ.1. F(w. In order to use the classifier’s margin to bound its expected loss. For 0-1 loss and linear metric ρL .1 189 A. Let γ1 > γ2 > . be a covering of a function class F(w. log2 N∞ (FL . .A.1. Let f : IR → [0.1. . from distribution D over i=1 X × Y. .5. m) = supS: |S|=m N∞ (F. where v(j) ∈ IRm . I(v Theorem A. if for all w there exists a v(j) such that for each data sample z(i) ∈ S: ρ(vi . z(i) )) ≤ . .1. be .1 Binary classification In binary classification. S) with -precision under the metric ρ.2 (Theorem 4 from Zhang [2002]) If w > 0. z)). where X = IRn and Y is mapped to ±1. V (j) is a covering of F(w.1. v(r) }. . the corresponding γ-margin loss is f γ (v) = 1 ≤ 2γ). we are given a sample S = {x(i) . Theorem A. . margin-based loss on the training sample that measures the worst loss achieved by the approximate covering based on this margin. S) = inf |V | s. then ∀ log2 (2 4ab/ + 2 m + 1) . ρL . f γ (v) = supρ(v.t. PROOF OF THEOREM 5. S). the bounds below use a stricter. Definition A.1 (Covering Number) Let V = {v(1) . . 1] be a loss function. S).v )<2γ f (v ). 1] be a loss function and f γ (v) = supρ(v. ρ. The covering number of a sample S is the size of the smallest covering: N∞ (F.3 (Corollary 1 from Zhang [2002]) Let f : IR → [0. We also define the covering number for any sample of size m: N∞ (F. In binary classification. m) ≤ 36 a2 b2 2 2 ≤ a and z 2 ≤ b. . y (i) }m . The next theorem bounds the expected f loss in terms of the γ-margin loss. . we have the following upper bound on the covering number of linear functions under the linear metric ρL (v. on the training sample. so we can fold x and y into z = yx. we let f (v) = 1 ≤ 0) be the step function. . When the norms of w and z are bounded. ρ. ρ.

y). . . z)) . In order for our loss function to compute average per-label loss.1. γi ≤ γ. γi . z). z))] ≤ ES [f γ (F(w. Definition A. which is a vector of minimum values of w ∆fi (y) for each error level H (y) from 1 to L as described below. z) = min w ∆fi (y). . . We will bound the corresponding covering number by building on the bound in Theorem A. . and pi be a sequence of positive numbers such that ∞ i=1 pi = 1. We define a new function class FM (w.190 APPENDIX A. We can now compute the average per-label loss from FM (w. with probability of at least 1 − δ over data: 32 1 ln 4N∞ (F.5 (Multi-error-level function class) The multi-error-level function class FM (w. z) = (M1 (w.2 Structured classification We will extend this framework to bound the average per-label loss H (y)/L for structured classification by defining an appropriate loss f and a function class F (as well as a metric ρ) such that f (F) computes average per-label loss and f γ (F) provides a suitable γ-margin loss.4 (dth-error-level function) The dth-error-level function Md (w. z). z))] + for all w and γ. ρ. PROOFS AND DERIVATIONS a decreasing sequence of parameters. . .2. z) is given by: FM (w. .t. z).1. ML (w. L} is given by: Md (w.1. it is convenient to make our function class vector-valued (instead of scalar-valued as above). . We can no longer simply fold x and y. A. y: H (y)=d Definition A. z) for d ∈ {1. S) + ln m pi δ ED [f (F(w. .1. . so we let z = (x. where for each fixed γ. since y is a vector. . z) by defining an appropriate loss function fM . we use i to denote the smallest index s. then for all δ > 0. Md (w. .

Since the margin of the hypothesis grows with the number of mistakes.v )≤2γ fM (v ). 1] is given by: fM (v) = 1 arg min vd . Otherwise. as there is more room for error.A.1. z)) = sup v:|vd −Md (w. d:vd ≤0 L where in case ∀d.6 (Average per-label loss) The average per-label loss fM : IR L → [0.z)|≤2dγ 1 arg min vd . z) > 0 corresponds to the classifier making no mistakes: arg maxy w fi (y) = y. d We now define the corresponding γ-margin loss using the new metric: Definition A. Combining the two definitions.1 191 Definition A. Md (w. y ).1. 1] is given by: γ fM (v) = sup ρM (v.7 (Multi-error-level metric) Let the multi-error-level metric ρM : IRL × IRL → IR for a vector in IR L be given by: ρM (v. PROOF OF THEOREM 5.1. v ) = max d |vd − vd | . z)) = 1 1 arg min Md (w. z) ≥ d:Md (w. d:vd ≤0 L . y Note that the case ∀d. we define arg mind:vd ≤0 vd ≡ 0. it is adversarial: it picks from all y which are better (w f (y) ≤ w f (y )). one that maximizes the Hamming distance from y. we get: γ fM (FM (w.8 (γ-margin average per-label loss) The γ-margin average per-label loss γ fM : IR L → [0.1. our metric can become “looser” with the number of mistakes.z)≤0 L L H arg max w fi (y) . With the above definitions.5. we have an upper bound on the average per-label loss fM (FM (w. We now need to define an appropriate metric ρ that in turn defines γ-margin loss for structured classification. This upper bound is tight if y = arg maxy w f (x. vd > 0. Definition A.

Vi .1. ρL . z(i) )) ≤ . S:|S|=m . ρL . . S ) for any sample S of size m. ρM . q. S).10 (Bound on multi-error-level covering number) N∞ (FM . m) = sup N∞ (FM . be a covering of FM (w. .9 (Multi-error-level covering number) Let V = {V(1) . . where V(j) = (V1 . . by definition. V(r) }. . . Consider a first-order sequence model as an example. m) ≤ N∞ (FL . S). ρM . ρL . . S) ≤ N∞ (FL . ρM . . S) ≤ N∞ (FL . . where we construct the sample S of size mNc (Vc − 1) in order to cover the clique potentials as described below. We also define N∞ (FM . ρM .1. Then Nc = 2L−1 since we have L node cliques and L − 1 edge cliques. FM (w. and V the number of values a variable takes. S). q is the maximum number of cliques that have a variable in common. V is a covering of FM (w. and q = 3 since nodes in the middle of the sequence participate in 3 cliques: previous-current edge clique. Vm ) and Vi sample z(i) ∈ S: ρM (Vi . Note that this is sufficient since N∞ (FL . ρL . The covering number of a sample S is the size of the smallest covering: N∞ (FM . . q. and current-next edge clique. . ρM . m) = sup N∞ (FM . and Rc is an upper-bound on the 2-norm of clique basis functions.t. so N∞ (FM . . . Lemma A. ρM . q. S ) ≤ N∞ (FL . . node clique. Recall that Nc is the maximum number of cliques in G(x). . . . S) = inf |V| s. with - precision under the metric ρM . . ρL . S: |S|=m (j) (j) (j) (j) (j) ∈ IRL . . with L as the maximum length. mNc (Vc − 1)). . Vc is the maximum number of values in a clique |Yc |.192 APPENDIX A. Vc = V 2 because of the edge cliques. PROOFS AND DERIVATIONS We also define the corresponding covering number for our vector-valued function class: Definition A. if for all w there exists a V(j) such that for each data We provide a bound on the covering number of our new function class in terms of a covering number for the linear function class. mNc (Vc − 1)). mNc (Vc − 1)). q. ρM . Proof: We will show that N∞ (FM .

∀c ∈ C (i) . Md (v(j) . . . at most dq vi. (j) (j) ∀i. We construct a set of vectors (j) (i) V = {v(1) . Vi . . . ∀y : H i (y) = d.2) (i) Note that vi. To make ∞-norm covering of FL (w. ∀yc . as ∆fi. . . . For each sample z ∈ S. . as each label can appear in at most q cliques. By combining this fact with Eq. and (j) Md (vi . one for each clique c and each assignment yc = yc . we require that for any w there exists a v such that for each sample z : |vi. we define vi. we create Nc (Vc − 1) samples ∆fi.c (yc ) = 0. .c (yc ) is zero for all cliques c for which the assignment is correct: yc = yc . Vm ) (j) (j) and Vi (j) ∈ IRL . .1. . .1 193 The construction of S below is inspired by the proof technique in Collins [2001].c (yc ). .c (yc ) = 0 for correct label assignments.3) .c (yc ). . This reduction in size comes about because our covering approximates the values of clique potentials w ∆fi. (j) (j) (j) (A. . S ) under ρL . where v(j) ∈ IRmNc (Vc −1) .c (yc )| ≤ . z(i) ) = miny: H (y)=d vi (y). ML (v(j) . . . . we obtain: vi (y) − w ∆fi (y) ≤ dq .c (yc ).A. .1) By Definition A. v(r) }. (A. . then i Vi = (M1 (v(j) . .1. z(i) )) . (j) (i) (j) (i) (i) V an (j) ∈V (A. We can now use V to construct a covering V = {V(1) . z(i) ). . . ρL . but the key difference is that our construction is linear in the number of cliques Nc and exponential in the number of label variables per clique. . The component of v(j) corresponding to the sample z(i) and the assignment yc to the labels of the clique c will be denoted by vi. mNc (Vc −1)). where (j) V(j) = (V1 . PROOF OF THEOREM 5. Let vi (y) = (j) (j) c vi. . while his is exponential in the total number of label variables per example. For convenience. (A. for our multi-error-level function FM . V(r) }.5.1). the number of vectors in V is given by r = N∞ (FL . z(i) ).1.c (yc ) for each clique c and clique assignment yc as opposed to the values of entire assignments w ∆fi (y).c (yc ) − w ∆fi. Thus for an assignment y with d mistakes.c (yc ) will be non-zero.

Proof: Substitute Theorem A. Coni i w v sider the case where vi (yd ) ≥ w ∆fi (yd ) (the reverse case is analogous). z)) ≤ ES fM (FM (w.4) v v w where the first step follows from definition of yd .1. Let yd = arg miny: H (y)=d vi (y) and yd = arg miny: H (y)=d w ∆fi (y). Theorem 5. z(i) )) = max d | miny: H (y)=d vi (y) − miny: H (y)=d w ∆fi (y)| i i d (j) (j) . (A. w v Fix any i. z(i) )) ≤ q . Lemma A.1 follows from above theorem with γi = Rc w 2 /2i and pi = 1/2i using an . (j) (j) (j) (j) (A. S) + ln m pi δ for all w and γ. PROOFS AND DERIVATIONS We conclude the proof by showing that V is a covering of FM under ρM : For each w. ρM . γ Theorem A. where for each fixed γ. ρM . with probability of at least 1 32 ln 4N∞ (FM . and pi be a sequence of positive numbers such that 1 − δ over data: γ Ez fM (FM (w. 2002] where in Step 3 (derandomization) we substitute the vector-valued FM and the metric ρM . be a decreasing sequence of parameters. Let γ1 > γ2 > . Proof: Similar to the proof of Zhang’s Theorem 2 and Corollary 1 [Zhang. We must now bound: (j) ρM (Vi . we use i to denote the smallest index s. we must (j) prove that: v w w w vi (yd ) − w ∆fi (yd ) ≤ vi (yd ) − w ∆fi (yd ) ≤ dq .2 into Lemma A. FM (w. The last step is a direct consequence of Eq.194 APPENDIX A.1.3). FM (w. z)) + ∞ i=1 pi = 1. .5.1. γi .10.1). (A. pick V(j) ∈ V such that the corresponding v(j) ∈ V satisfies the condition in Eq. .3) Let fM and fM (v) be as defined above. then for all δ > 0.12 (Multi-label analog of Theorem A. .1. m) ≤ 36 2 Rc w 2 2 2 (j) q2 log2 1 + 2 4 Rc w 2 q + 2 mNc (Vc − 1) .11 (Numeric bound on multi-error-level covering number) log2 N∞ (FM . Hence ρM (Vi . γi ≤ γ. since vi (yd ) ≤ vi (yd ).t.1.

. λ(k) = minv : µv (k)>0 yv (k) for k = 1. we can apply it repeatedly to get an optimal integer solution.1) Consider any fractional.5) s.A.2 AMN proofs and derivations In this appendix. µv (k) ≥ 0 and µc (k) = mini∈c µv (k). I(0 Now consider the smallest fractional µv (k). Recall that the LP relaxation for finding the optimal maxy g(y) is: K K max v∈V k=1 µv (k)gv (k) + c∈C\V k=1 K µc (k)gc (k) µv (k) = 1. Since gc (k) ≥ 0. AMN PROOFS AND DERIVATIONS 195 argument identical to the proof of Theorem 6 in Zhang [2002].2.2. ∀c ∈ C \ V. k. µc (k) ≥ 0. To show that µ is feasible. otherwise we could increase the objective by increasing µc (k). 2. we can assume that µc (k) = minv∈c µv (k). ∀c ∈ C. µc (k) ≤ µv (k). we present proofs of the LP inference properties and derivations of the factored primal and dual max-margin formulation from Ch. k=1 (A. I(0 µc (2) = µc (2) + λ1 < µc (2) < 1). I(0 µv (2) = µv (2) + λ1 < µv (2) < 1).t. µ will have at least one more integral µv (k) than µ. Note that if λ = λ(1) or λ = −λ(2). feasible µ. v ∈ c.2. ∀v ∈ V. 7.1 Binary AMNs Proof (For Theorem 7. k. A. We construct an assignment µ from µ by leaving integral values unchanged and uniformly shifting fractional values by λ: µv (1) = µv (1) − λ1 < µv (1) < 1). A. I(0 µc (1) = µc (1) − λ1 < µc (1) < 1). We show that we can construct a new feasible assignment µ which increases the objective (or leaves it unchanged) and furthermore has fewer fractional entries. we need µv (1) + µv (2) = 1. Thus if we can show that the update results in a feasible and better scoring assignment.

To show that µv (k) ≥ 0. and it remains to show that we can improve the objective. we show that µv (1) + µv (2) = 1.196 APPENDIX A. I(0 Hence the new assignment µ is feasible. does not decrease the objective function.2   = λ v∈V [Dv (1) − Dv (2)] + c∈C\V [Dc (1) − Dc (2)] = λD Dc (k) = gc (k)1 < µc (k) < 1).2 [µc (k) − µc (k)]gc (k) c∈C\V k=1. will necessarily increase the objective (or leave it unchanged). we prove minv µv (k) = 0. The change in the objective is: [µv (k) − µv (k)]gv (k) + v∈V k=1. PROOFS AND DERIVATIONS First. since µv (1) + µv (2) = 1. I(0 i∈c i∈c i∈c µc (2) = µc (2) + λ1 < µc (1) < 1) I(0 = (min µv (2)) + λ1 < min µv (2) < 1) = min µv (2). Lastly. We can show that the change in the objective is always λD for some constant D that depends only on µ and g. This implies that one of the two cases. min i i:µv (k)>0 µv (k) − min µv (k) i:µv (k)>0 = 0. and . so is µv (2). we show µc (k) = mini∈c µv (k). µc (1) = µc (1) − λ1 < µc (1) < 1) I(0 = (min µv (1)) − λ1 < min µv (1) < 1) = min µv (1). µv (1) + µv (2) = µv (1) − λ1 < µv (1) < 1) + µv (2) + λ1 < µv (2) < 1) I(0 I(0 = µv (1) + µv (2) = 1. I(0 i∈c i∈c i∈c We have established that the new µ are feasible. λ = λ(1) or λ = −λ(2). min µv (k) = min µv (k) − ( min µv (k))1 < µv (k) < 1) I(0 v v i:µv (k)>0 = min min µv (k). I(0 Dv (k) = gv (k)1 < µv (k) < 1). Above we used the fact that if µv (1) is fractional.

2 Multi-class AMNs For K > 2. the probability that a node is as- signed label k over all phases is exactly µv (k). Nodes in the clique c will be assigned label k by the procedure if they are assigned label k in one phase. and a threshold α ∈ [0. For each node i which has not yet been assigned a label. we use the randomized rounding procedure of Kleinberg and Tardos [1999] to produce an integer solution for the linear relaxation. The procedure terminates when all nodes have been assigned a label. Lemma A.2. By symmetry.2. (They can also be assigned label k as a result of several phases. AMN PROOFS AND DERIVATIONS 197 has strictly fewer fractional entries.A.) The probability that all the . |c| c Proof For a single phase. losing at most a factor of m = maxc∈C |c| in the objective function.2. In each phase. uniformly at random. the probability that all nodes in a clique c are assigned label k if none of the nodes were previously assigned is 1 K mini∈c µv (k) = that at least one of the nodes will be assigned label k in a phase 1− 1 K K k=1 1 µ (k). K v which is proportional to µv (k). Proof The probability that an unassigned node is assigned label k during one phase is 1 µ (k).2 The probability that all nodes in a clique c are assigned label k by the procedure is at least 1 µ (k). Our analysis closely follows that of Kleinberg and Tardos [1999]. Lemma A.1 The probability that a node i is assigned label k by the randomized procedure is µv (k). The probability that none of the nodes in the clique will be assigned any label in one phase is maxi∈c µv (k). we assign the label k if µv (k) ≥ α. A. we pick a label k.2. 1] uniformly at random. but we can ignore this possibility for the purposes of the lower bound. The probability K c 1 is K (maxi∈c µv (k)). The basic idea of the rounding procedure is to treat µv (k) as probabilities and assign labels according to these probabilities in phases.

. in fact. PROOFS AND DERIVATIONS nodes in c will be assigned label k by the procedure in a single phase is: ∞ j=1 1 1 µc (k) 1 − K K µc (k) K k=1 i∈c K j−1 max µv (k) k=1 i∈c = = µc (k) K k=1 maxi∈c µv (k) ≥ µv (k) = i∈c µc (k) K k=1 µv (k) ∞ i=0 µc (k) . c Proof This is immediate from the previous two lemmas. (A.3 The expected cost of the assignment found by the randomized procedure given a solution µ to the linear program in Eq. We can also derandomize this procedure to get a deterministic algorithm with the same guarantees. using the method of conditional probabilities. The only difference between the expected cost of the rounded solution and the (non1 integer) optimal solution is the |c| factor in the second term. similar in spirit to the approach of Kleinberg and Tardos [1999]. As node potentials are often larger in magnitude than clique potentials.5) is at least 1 c∈C\V |c| K k=1 v∈V K k=1 gv (k)µv (k) + gc (k)µk . we observe.198 APPENDIX A.5). only to the clique potentials. 1−d Above. the potentials associated with large cliques are typically smaller in magnitude. Note that the approximation factor of m applies. Thus. the max of the set of positive µv (k)’s by their sum. if we compare the log-probability of the optimal MAP solution and the logprobability of the assignment produced by this randomized rounding procedure. di = and then upper-bounded Theorem A. We obtain an additive error (in log-probability space) only for the clique potentials. we note that the constant factor approximation is smaller for smaller cliques. again. the fact that we incur no loss proportional to node potentials is likely to lead to smaller errors in practice. Along similar lines. reducing further the actual error in practice. |c| 1 . (A. we first used the fact that for d < 1. the terms corresponding to the log-partition-function and the node potentials are identical. By picking m = maxc∈C |c|. we have that the rounded solution is at most m times worse than the optimal solution produced by the LP of Eq.2.

(i) (i) ¨ f Now let ξi.v (k) − c⊃v ∀i.7) ξi. k.c (k). .c. v∈V (i) w fi (y ) + ξi ≥ −w fi.3 Derivation of the factored primal and dual max-margin QP Using Assumptions 7.A. ∀i.c (k.v (yv ) + c⊃v w ¨i. mi.v (k) for each of the Substituting this dual into Eq. we have the dual of the LP used to represent the interior max subproblem maxy w fi (y) + i (y) in Eq. (7. AMN PROOFS AND DERIVATIONS 199 A. we obtain: min s. c ∈ C (i) \ V (i) .t.v for each normalization constraint in Eq.2.c.4.c. ∀i.1) and variables mi.v . Re-expressing the above QP in terms of these new variables.v (k) i.1 and 7.2): min v∈V ξi. v ∈ V (i) . . ¨ f −w ¨i. k). k. k. ∀i. .c.c.v + w fi.v (k) ≥ mi.v .c (yc )/|c|.v (k) i. k. ¨ f −w ¨i. v ∈ c.c (k) = fi. .c (k) + mi. s. (5.v (k) + (i) ¨ f w ¨i.v (k) ≥ 0. . . ∀i. c ∈ C (i) \ V (i) . ¨ w ≥ 0. . (3. v ∈ c. i.t.2.v = ξi.c.c. where fi. k.c (k) = i. k) and inequality constraints.1).c.v (k) ≥ 0.c (k.c. − ξi. v ∈ V (i) . c ∈ C (i) \ V (i) . c ∈ C (i) \ V (i) . we get: .6) mi.c (k) + mi.v .4. 1 ||w||2 + C 2 (i) ξi i (A.c (yc )/|c| and mi. . In the dual. we have a variable ξi.v (k) ≥ mi.v (k) = mi.c (k). ∀i.v −w fi.v (k) − c⊃v (A.4. i. k.v (k) ≥ v∈c i. ∀i.v (k) ≥ v∈c − ξi.

c (yc )/|c|.c.v∈V (i) (A.v (k) − ¨ w ∆¨i.c. we can eliminate ξi and the corresponding set of constraints to get the formulation in Eq. ∀i. (i) ¨ f mi. repeated here for reference: min s.c (k) + f v∈c − ξi. mi. c ∈ C (i) \ V (i) .c. c ∈ C (i) \ V (i) .v .c (k).v (k) ≥ c⊃v i. v ∈ c. (7.v (k) − c⊃v mi. ¨ w ≥ 0. (i) ¨ f mi.c (k) + f ∀i. ¨ w ≥ 0.v (k) i.v . ∀i. c ∈ C (i) \ V (i) .v (k) ≥ −w ¨i. k. w ∆fi. v ∈ c.8) ξi.9) i. PROOFS AND DERIVATIONS min s. k. k. k. Since ξi = i. c ∈ C (i) \ V (i) .t. ∀i. k. 1 ||w||2 + C 2 ξi ≥ v∈V (i) ξi i (A.c. w ∆fi.4). ∀i. v ∈ V (i) .c.c (yc )/|c|.v at the optimum. mi.v (k) i.v (k) ≥ mi.c (k).v (k) ≥ v∈c − ξi. ∀i.v i.c.200 APPENDIX A.t. ¨ w ∆¨i. v ∈ V (i) . .v∈V (i) ξi. k. 1 ||w||2 + C 2 ξi.v (k) ≥ −w ¨i.v .v (k) ≥ ∀i.

c∈C (i) . k (A.9) is given by: 2 max i. k.v (k) = λi. v ∈ c.c.c (yc ) ≥ 0. ¨ .c. ¨ ∀i.c (k) ≤ µi. µi. ∀v ∈ V (i) . ν ≥ 0. ∀i.v (k)∆fi. µi.v∈c.v (k)¨i.v∈V (i) . \ V . while λ and τ correspond ¨ to third and fourth set of constraints. and re-express the above QP as: 2 2 max i.t. we can eliminate λ and τ .c (k) − µi.c (k) f i. ∀c ∈ C (i) . µ correspond to the first two sets of constraints.c (k) f i. v ∈ c.c (k) i. ∀c ∈ C (i) \ V (i) . k 1 µi.v∈V (i) .c∈C (i) . ∀c ∈ C .c. k. ∀i. ∀i. Using the substitution ν=τ+ ¨ ¨ i. ∀c ∈ C (i) \ V (i) . µi. ∀c ∈ C (i) k=1 (i) µi.v (k) ≥ 0 τ ≥ 0.v (k) = 1.c∈C (i) \V (i) .v (k).v∈c. λi. k=1 µi.c (yc )/|c| f ¨ and the fact that λi. k (i) i.c (k) ≥ 0. v ∈ c. k s. k.10) 2 (i) λi.c (k) ≥ 0.v i.c∈C (i) .A.v (k)¨i.v (k) = C.v (k). k s.v i.c.c (k) − 2 1 τ+ ¨ − 2 ˙ µi.t. k (i) λi. k.c (k) i.v (k) ≥ 0 and ¨i.c (k)∆¨i. ∀v ∈ V (i) . as well as divide f (i) µ’s by C. µi. AMN PROOFS AND DERIVATIONS 201 Now the dual of Eq. k. k K 1 ¨ − C ν+ 2 µi.c∈C (i) \V (i) . k 1 µi. ∀i. (A. ∀i.c (k) − C 2 ˙ µi.v (k)∆fi.2.c∈C (i) .c (k)∆¨i.c (yc )/|c| + f K µi. In this dual.c. ∀i.

A. T. (2003). & Ng. Proc. F. & Hofmann. Exponential families for conditional random fields. T. Tsochantaridis.. & Orlin. Madden. Altun. F. A. (2003). A. P. Bach. S. (2004). ICML. Bach. Altschul. NIPS. I.. Y.. AddisonWesley-Longman. Large-scale prediction of disulphide bond connectivity... & Apweiler. & Ribeiro-Neto. Y. Apprenticeship learning via inverse reinforcement learning. Modern information retrieval. 35. P. ICML. & Hofmann. Proc. R. 3389–3402. Inverse optimization. Hidden markov support vector machines. 202 . UAI. (2004). Operations Research.. Smola. Proc. (2004). & Vullo... (2001). Learning spectral clustering. & Lipman (1997)..... Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. NIPS.. Miller. B. A. Bairoch.. & Jordan. A. 45–48.. M. Schaffer. T. Proc. 28. J. W. Thin junction trees. Nucleic Acids Res. Part I: Linear programming and general problem. Nucleid Acids Res. NIPS.Bibliography Abbeel. Cheng. & Jordan. (2001).. R... Y. 25. Baeza-Yates. Altun. Zhang. J. The Swiss-Prot protein sequence database and its supplement trembl. Ahuja. M.. Baldi. 771–783. (1999). A. (2000). R.

IEEE Trans. A. Malik. Bishop. The protein data bank. 24. Blum. 48. E. & Puzicha. Pattern Anal.. L. Taskar. FOCS. Correlation clustering. and risk bounds (Technical Report 638). Belmont.. & Shapiro. N. Athena Scientific. Combining labeled and unlabeled data with co-training. Berman. (2004). Bartlett. Intell.. J. Feng. Shindyalov.. ICML. (2003).. Collins. (1995).. A. M. (1999). Introduction to linear programming. Shape matching and object recognition using shape contexts. & Tsitsiklis. Proc. The principled design of large-scale recursive neural network architectures dag-rnns and the protein structure prediction problem. Jordan. Convexity. (1994). Z. 575–602. Gilliland. (2003). H. D. P.. Belongie. S. T. Nonlinear programming. Besag. Bhat. M. C. classification. (2002). A. Journal of the Royal Statistical Society B. & McAuliffe..... Bansal. P.. G. & Bourne. On the statistical analysis of dirty pictures.. Exponentiated gradient algorithms for large-margin structured classification.. H. MA: Athena Scientific. Proc. N. D. I. (2002). Berkeley. & McAllester. T. Shental. 4. I. Bertsekas. 3(2). Journal of Machine Learning Research. P. P. & Chawla. Bartlett.. & Mitchell.. D.. Westbrook. Bouman. Oxford. COLT. J. Department of Statistics. G. (1997). U.. Bertsimas.C.BIBLIOGRAPHY 203 Baldi. IP.. Bar-Hillel. UK: Oxford University Press. Mach... (2003). J. (1998). (2000). Neural networks for pattern recognition.. & Weinshall. Weissig. . C. & Pollastri. J. J. T. Blum. Learning distance functions using equivalence relations.. M. D. S. (1986). Hertz. J.. D. A multiscale random field model for bayesian image segmentation. NIPS. B.

J. S. 53. & Schoelkopf. Y. Weston. (2004). A. & Zabih. Head-driven statistical models for natural language parsing. Burton.. (1999b). FOCS. SIGMOD. (1992).. A. Convex optimization. & Toint. M. O.. 35.. PAMI. NIPS.. B. S. Journal of VLSI Signal Processing. D. On an instance of the inverse shortest paths problem. R. (2003). ICCV. 287–295. (1998). P. & Curran. Clark. R. Predicting the disulfide bonding state of cysteines with combinations of kernel machines.. P. & Indyk.. (2003). Passerini. Statistical language learning. Boykov. Y. (1993).. V. & Vullo. (2002). Boykov. & Zabih.. Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL ’04). Frasconi. University of Pennsylvania. Proc. Parsing the wsj using ccg and log-linear models. O. Charniak.. Guruswami. Collins. & Kolmogorov. Cambridge University Press. P. Markov random fields with efficient approximations. R. Cluster kernels for semi-supervised learning.. L. & Wirth..204 BIBLIOGRAPHY Boyd. Mathematical Programming. E. Charikar. An experimental comparison of min-cut/max-flow algorithms for energy minimization in computer vision. CVPR.. Veksler. (1999a). A. L. 45–61. Enhanced hypertext categorization using hyperlinks. B. S. V. O. . Chakrabarti. (2004). Doctoral dissertation. & Vandenberghe. Boykov.. A. Y. M... Fast approximate energy minimization via graph cuts. Dom. J. (1999). MIT Press.. Clustering with qualitative information. Veksler. Ceroni. Chapelle. (2004).

Leiserson. Della Pietra. Mitchell. (1970). D. An introduction to support vector machines and other kernel-based learning methods. (2000). A... Bunt. (2003).. Crammer. D. 19(4). J. J. Learning to extract symbolic knowledge from the world wide web. S. Cover.). (1999). Della Pietra.. S.. M.. IEEE Trans.. & Singer. T. R. APPROX. Lauritzen. M. Nigam. New developments in parsing technology. on Pattern Analysis and Machine Intelligence. 380–393. Inducing features of random fields. In H. Collins. Correlation clustering with partial information. 175–182). & Jaakkola. Discriminative reranking for natural language parsing. Parameter estimation for statistical parsing models: Theory and practice of distribution-free methods. R. (1997)... Rivest. Cowell.. New York: Springer. C. T. Elements of information theory. K.. New York: McGraw-Hill. K.. Introduction to algorithms. Y. Proc. Dawid. M. UAI. D. Corduneanu. Probabilistic networks and expert systems. (2000). M. 2nd edition. Satta (Eds. IWPT.. Carroll and G. & Lafferty. 509–516). A. 265–292. A. . New York: Wiley. (1991). H. Journal of Machine Learning Research. 2(5). A. DeGroot. V. & Spiegelhalter. (2004). DiPasquo. & Thomas. & Stein... D. Craven. C.. (2001). T. Proc AAAI98 (pp. T.. Cambridge University Press. N. & Slattery. Optimal statistical decisions. Demaine. Collins. M. (1998). J. McCallum. Freitag.. M. ICML 17 (pp. (2001). On information regularization. S. & Immorlica. J.BIBLIOGRAPHY 205 Collins.. Cormen. N. (2001). Kluwer. (2003). E. MIT Press. On the algorithmic implementation of multiclass kernelbased vector machines. Parameter estimation for statistical parsing models: Theory and practice of distribution-free methods. & Shawe-Taylor. Cristianini.

... D. 3. P. . A. D.. & Mitchison. Hart. & Fiat. G.. D. S. 957–964. 17. (1977). Bionformatics. & Sperduti. Computer vision: A modern approach. 340–346. Pattern classification. ESA. 39. P. (1990). Bioinformatics. & Ponce. 251–256. G.. IEEE Transactions on Neural Networks. (1998).. L. O. Journal of Research at the National Bureau of Standards... P. (1998). Probabilistic theory of pattern recognition. Correlation clustering . A neural network-based method for predicting the disulfide connectivity in proteins. Introduction to informetrics. J. P. Proteins. R. (1965). A general framework for adaptive structures. & Rubin. M.minimizing disagreements on arbitrary weighted graphs. Ricobelli. Frasconi. 1. & Rousseau. Fariselli.206 BIBLIOGRAPHY Dempster. R. Fiser. Predicting the oxidation state of cysteines by multiple sequence alignment. Gy¨ rfi.. A. B. M. Devroye. G.. 69B. A. Journal of the Royal Statistical Society. Maximum likelihood from incomplete data via the EM algorithm. & Casadio. & Lugosi. Laird. 464–468. (2000). (1999). Biological sequence analysis. R. Prediction of disulfide connectivity in proteins. Elsevier. P. Prentice Hall.. Duda.. J.. P. & Simon. 2nd edition. P. Krogh. E. Maximum matching and a polyhedron with 0-1 vertices. & Casadio. Emanuel. & Stork. Fariselli.. 125–130. R. Durbin.. R. Egghe. Martelli. (2002). Role of evolutionary information in predicting the disulfide-bonding state of cysteine in proteins. & Casadio. R. Cambridge University Press.. N. (2000).. o New York: Springer-Verlag. P. New York: Wiley Interscience. (2001). A.. Eddy. L. Fariselli. (2003).. (1996). 36. Knowledge based intelligent information engineering systems and allied technologies (KES 2002). L. D. Edmonds. Gori. A. (2002). I. Forsyth.

& Venkataraman. Gusfield. L. Journal of Machine Learning Research. Wash. 19. Efficient solution algorithms for factored mdps. Stanford University.. Geman. Probabilistic models of text and link structure for hypertext classification.. Sweden. Parr. N. N. & Campbell. R. P. H.. IJCAI99 (pp.. 51. Guestrin. Large margin classification using the perceptron algorithm. Proc. Proc. T... Getoor. Greig. L. Friess. Exact maximum a posteriori estimation for binar images. (2001).. Computational Learing Theory. Koller. D. Friedman. E. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. D. The kernel adatron algorithm: a fast and simple learning procedure for support vector machine. Proceedings of IEEE Neural Network for signal processing conference. Freund. & Vullo. Implementation of algorithms for maximum matching on nonbipartite graphs. Getoor. D. R.. M. & Schapire. and sequences : Computer science and computational biology.. (2002). & Taskar.. A two stage svm architecture for predicting the disulfide bonding state of cysteines. ICML. L. Cristianini. D.. Stockholm. A. A. & Seheult. T.. Cambridge University Press.. C... Porteous. M. D. (1999). & Pfeffer. S. Segal. (2002).. R. Koller. S. Getoor. Taskar. & Johnson. Y. (1997). Passerini. Proc. Computers and intractability. C. Garey. Statist. A. (1989).. M. B. (1998). & Koller.. Soc.. Seattle. Journal of Artificial Intelligence Research. B. trees. & Johnson. Algorithms on strings. Learning probabilistic models of link structure. Friedman. D. B. IJCAI01 Workshop on Text Learning: Beyond Supervision. J. Dynamic programming for parsing and estimation of stochastic unification-based grammars. 1300–1309). B. D. (1979). 8. R. Gabow. Freeman. Koller. S.. (1998). H. (1973). N. A. Doctoral dissertation..BIBLIOGRAPHY 207 Frasconi. (2002). (2003). Learning probabilistic relational models. .

& Riezler. Doctoral dissertation. Heuberger. King. & Crouch. Transductive inference for text classification using support vector machines. Hastie. J. A comparison of approaches to on-line handwritten character recognition. S. Proceedings of ACL 1999. MIT Spoken Language Systems Group. Convolution kernels on discrete structures (Technical Report)... & Sternberg. Exponentiated gradient versus gradient descent for linear predictors. (2001). T.). M. (1983). Analysis and classification of disulphide connectivity in proteins. Johnson. Kassel. Joachims. Kivinen. R. methods. Chi. Geman. J. & Friedman... . UC Santa Cruz. Morgan Kaufmann Publishers. 1–63. (1997). Kabsch. and results. D. A. (2004). Vasserman.. T. D. & Warmuth. US. (1999). 200–209). Johnson. Inverse combinatorial optimization: A survey on problems. Speed and accuracy in shallow and deep stochastic parsing.. M. R. S. Estimators for stochastic “unification-based” grammars. Tibshirani. Riezler. Canon. W. Proc.. M. R.. Haussler. Z. S. the entropic effect of cross-linkage. San Francisco. S. T. C. R. Maxwell. ACL 39.... J. (1999). Dictionary of protein secondary structure: Pattern recognition of hydrogen-bonded and geometrical features.. Hochbaum. 244. (2001). PWS Publishing Company. Journal of Combinatorial Optimization. (1999). Information and Computation. 132(1). S. M.208 BIBLIOGRAPHY Harrison. (Ed. (1997). Joint and conditional estimation of tagging and parsing models. (2004). C. & Sander. ICML99 (pp. Journal of Molecular Biology.. New York: Springer-Verlag. The elements of statistical learning.. Proceedings of HLT-NAACL’04. 8. P. (1995). Kaplan. Approximation algorithms for NP-hard problems. (1994).

Kleinberg. Z. (2003). Discriminative fields for modeling spatial dependencies in natural images. 423– 430). (2002).. FOCS. Kleinberg. & Liu. C. Proc. ICML. Accurate unlexicalized parsing. D. Eskin. Approximation algorithms for classification problems with pairwise relationships: Metric labeling and Markov random fields. 191–208..BIBLIOGRAPHY 209 Klein. E. & Noble. (2004).. Prediction of β-sheet topology and disulfide bridges in polypeptides. Authoritative sources in a hyperlinked environment. Wisc. NIPS. Kernel conditional random fields: Representation and clique selection. (2003). Lafferty. & Floudas. Lafferty. J. 46(5). Y. D.. AAAI98 (pp. & Hebert. Kolmogorov. C. Klepeis. Lawler. (2002). J.. Combinatorial optimization: Networks and matroids. S. Journal of Computational Chemistry. J. E.. Koller. On inverse problems of optimum perfect matching. (1976). Rinehart and Winston. Leslie. & Pfeffer. . Zhu. X. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. (1998). ACL 41 (pp. A. R.. New York: Holt. Journal of the ACM. F. M. 7. & Zhang. (1999). D. Liu. M. Pacific Symposium on Biocomputing.. (2003). & Tardos. (2001). (2003)... J. A. & Manning. 580–587).. V. W. McCallum. 24. What energy functions can be minimized using graph cuts? PAMI. J. ICML. Proc. & Pereira. J. Kumar. & Zabih. (1999).. 604–632.. Probabilistic frame-based systems. C. E. The spectrum kernel: a string kernel for svm protein classification. Madison. Journal of Combinatorial Optimization.

. ICCV.210 BIBLIOGRAPHY Lodhi. Toward conditional models of identity uncertainty with application to proper noun coreference... (1999). Text classification using string kernels. Proc. Computational Linguistics. (1993). P. R.. D. M.. NIPS.. N. D. L. 313–330. (2000). & Sch¨ tze. P. 27359. & Ney. Prediction of the disulfidebonding state of cysteines in proteins at 88% accuracy. Oxford University Press. & Wellner. 19(2). 11. (1997). Proc. R. Proc. & Casadio. P. Signor. Basil Blackwell. A. & Malik. Marcus. Mitchell. T. Symmetric word alignments for statistical machine translation. Investment science.. Matusov. C.. (2003). Machine learning. Substantial increase of protein stability by multiple disulfide bonds. Shawe-Taylor. C. Cristianini. Nature. (2004). McGraw-Hill. M. (2003). Matsumura. Santorini. Luenberger. & Watkins. COLING. Portfolio selection: Efficient diversification of investments.. Martin. & Matthews. Protein Science. B. Fowlkes. H. B. Zens. Markowitz. D. Building a large annotated corpus of English: The Penn Treebank.. H. E. u Cambridge. 342. Massachusetts: The MIT Press. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. M. & Marcinkiewicz. IJCAI Workshop on Information Integration on the Web. C. Foundations of statistical natural language processing. UAI. B. H. Efficiently inducing features of conditional random fields. Martelli. McCallum. . Malaguti. (1997).. (1989).. McCallum. J. 291:293. H. J. Manning.. (1991).. A.. (2001).. (2002). A.. G. Fariselli. Tal.

B. Text classification from labeled and unlabeled documents using em. Nocedal. Needleman. (1999). Journal of Molecular Biology. Proceedings of Human Language Technology Conference (HLT 2002). J. (1982). Papadimitriou. L. & Jordan.. New York: Springer. Pearl. K. Table extraction using conditional random fields. M. & Steiglitz. (2000).. Numerical optimization. T. K... Proc. A. C. & Wunsch. Pinto. S. Y. D. Loopy belief propagation for approximate inference: an empirical study. Proc. C. D. & Wright... McCallum. A. ACM SIGIR. & Croft. & Mitchell. S... Y. J. NJ: Prentice-Hall. Nigam. Englewood Cliffs. L. Weiss. (1999). Thrun. Proc. Iterative classification in relational data. Machine Learning. (2001). Wei. On discriminative vs. W. A general method applicable to the search for similarities in the amino acid sequences of two proteins. A. K. Murphy.BIBLIOGRAPHY 211 Miyao.. Ng. 467–475). Algorithms for inverse reinforcement learning. A. I. AAAI-2000 Workshop on Learning Statistical Models from Relational Data (pp.. UAI99 (pp. (2000). (1970). Y. Neville. S. NIPS. X. Ng. (2000). & Wolsey. (2002).. 13–20). P. J.. 48. Proc. McCallum. & Russell. Combinatorial optimization: Algorithms and complexity. (1988). Probabilistic reasoning in intelligent systems... San Francisco: Morgan Kaufmann.. & Jordan. J. & Tsujii. New York: J. generative classifiers: A comparison of logistic regression and naive Bayes. & Jensen. M. .. (1999). J. Maximum entropy estimation for feature forests. Wiley. Integer and combinatorial optimization. Nemhauser. S. 39. (2003). A. G. ICML.

Shen. Amsterdam: Elsevier. Springer.. Sarkar. & Joshi.. 195–197. ICML. Bartlett. & Mitchell. A. J.. (2001). Sha. C. Rohanimanesh. Proc. Cambridge Phil. Schrijver. Induction of decision trees. & McCallum. F. Inverse problem theory: Methods for data fitting and model parameter estimation. ICML00 (pp. NIPS. HLTNAACL. T. IEEE Trans. on Information Theory. 44(5). M.. EMNLP. (1998). Using sparseness and analytic QP to speed training of support vector machines. Dynamic conditional random fields: Factorized probabilistic models for labeling and segmenting sequence data. F. (1999). M. Shallow parsing with conditional random fields. NIPS. & Anthony. C. (2004).. A. K. 1926–1940. (1987). (1952). Szummer. Williamson. Discovering test set regularities in relational domains. Combinatorial optimization: Polyhedra and efficiency. (2003). M. Soc. A. Shawe-Taylor. (1981). (2003). Quinlan. 48. & Pereira. Biol. 1.. 895–902). Structural risk minimization over data-dependent hierarchies.. R. Some generalized order-disorder transformations. Machine Learning. L. Slattery. 147. Proc. & Waterman. Tarantola. Partially labeled classification with Markov random walks. T. R. L.212 BIBLIOGRAPHY Platt. Using ltag based features in parse reranking.. Proc. J. Identification of common molecular subsequences. Sutton. K.. & Jaakkola.. S. T. Potts. P. Proc. Proc. Mol. A. 81–106. A. Smith.. J. B. (2003). J.. .. R. Proc. (2001). (2000).

(1979). B. Taskar. New York: Springer-Verlag.. A. Max margin parsing. A. Tsochantaridis. D. A. New York. E.. Toutanova. (2003). Theoretical Computer Science.. Global protein function prediction from protein-protein interaction networksh. Max margin Markov networks. . (2003b). Proc. P. Wash. & Palsson.. & Koller.. B. K. 189–201. A.. Klein. Feature-rich part-of-speech tagging with a cyclic dependency network. Maritan. Taskar. Flammini. Hofmann. Support vector machine learning for interdependent and structured output spaces.. V.. 870–876). Taskar. (2004)... D. P. Abbeel. T. Valiant. C.. scientific and practical use. B... Y. B. (2003a).. (2002). Guestrin. I. Link prediction in relational data. NIPS. B. NIPS. Taskar.. Varma. & Koller. & Koller. Vapnik. Manning. UAI. 6. Discriminative probabilistic models for relational data. Nature Biotechnology... Collins. & Vespignani. & Koller.. Chatalbashev.. (2001). 8. Wong. D. & Manning. (2004a). Abbeel. Taskar.. (2004b). D.. D. Klein. L. (1995).BIBLIOGRAPHY 213 Taskar. NAACL 3 (pp. D. Koller. & Altun. Proc. The complexity of computing the permanent. (2003). Joachims. Bio/Technology. D. & Koller. D. M. C. ICML. & Singer. B. Seattle. Segal. Learning associative Markov networks. 252–259). Twenty-first international conference on Machine learning.. M. Probabilistic classification and clustering in relational data. 12. C. Metabolic flux balancing: Basic concepts. Vazquez.. B. G.. (1994). V. A. The nature of statistical learning theory. T. D. EMNLP. Y. Proc. IJCAI01 (pp..

E. Journal of Intelligent Information Systems. & Frasconi... risk estimation. Allerton Conference on Communication.. Map estimation via agreement on (hyper)trees: Message-passing and linear programming approaches. Zhang. (1993). & Russell. Xing. J. A. & Watkins. a. Optimization. Younger. (2002). . J. Weston. M. via penalized log likelihood and smoothing spline analysis of variance. A. H. University of London. Generalized belief propagation. Y. 59–72. Control and Computing. Cornell University. I. Jaakkola.. 653–659. R. Computational Learning Theory and Natural Learning Systems. M. & Chappell. Wainwright. D. Variational inference in graphical models: The view from the marginal polytope. S.a. Y. R. (1998). Bioinformatics. Proc. A study of approaches to hypertext categorization. (2002). & Weiss.. (2000). Distance metric learning. 20. Vullo. Wang. 189–208. Information and Control. Efficient graph-based energy minimization methods in computer vision. Disulfide connectivity prediction using recursive neural networks and evolutionary information. Department of Computer Science. Royal Holloway. Recognition and parsing of context-free languages in time n3 . & Ma. (2002). C.214 BIBLIOGRAPHY Veksler. Slattery. Soft classification. & Jordan. P.. Wainwright. Yang. 18(2). NIPS. S. with application to clustering with side-information. Z. J. Allerton Conference on Communication. A network flow method for solving inverse combinatorial optimization problems. (1999). Yedidia.. A. M. Proc. 37. (1967). Freeman. NIPS... 10.. Gu. Ng. W. & Willsky. (2004). (1996).. Control and Computing.. & Ghani. M.. Wahba. Multi-class support vector machines (Technical Report CSD-TR-98-04). C. T.. (2003). Doctoral dissertation. Jordan.. O. Y.k.. G.

BIBLIOGRAPHY 215 Zhang.. (2002). ICML. . NIPS. & Hastie. 2. Zhu. Proc. Kernel logistic regression and the import vector machine. Ghahramani. 527–550. (2003). Journal of Machine Learning Research. T. J. T. Zhu. Covering number bounds of certain regularized linear function classes. Proc.. (2001). & Lafferty. J.. X. Z. Semi-supervised learning using gaussian fields and harmonic functions.