A Clusterng K-Means VCLASS 4

Data Mining What is Cluster Analysis?
Cluster Analysis: Basic Concepts

and Algorithms z Finding groups of objects such that the objects in a group
will be similar (or related) to one another and different
from (or unrelated to) the objects in other groups
Lecture Notes for Chapter 8
Inter-cluster
Intra-cluster distances are
Introduction to Data Mining distances are maximized
minimized
by
Tan, Steinbach, Kumar
© Tan,Steinbach, Kumar Introduction to Data Mining 4/18/2004 1 © Tan,Steinbach, Kumar Introduction to Data Mining 4/18/2004 2
Applications of Cluster Analysis What is not Cluster Analysis?

Discovered Clusters Industry Group
z Understanding Applied-Matl-DOWN,Bay-Network-Down,3-COM-DOWN, z Supervised classification
– Group related documents
1 Cabletron-Sys-DOWN,CISCO-DOWN,HP-DOWN,
DSC-Comm-DOWN,INTEL-DOWN,LSI-Logic-DOWN,
Micron-Tech-DOWN,Texas-Inst-Down,Tellabs-Inc-Down,
Technology1-DOWN – Have class label information
Natl-Semiconduct-DOWN,Oracl-DOWN,SGI-DOWN,
for browsing, group genes Sun-DOWN
Apple-Comp-DOWN,Autodesk-DOWN,DEC-DOWN,
and proteins that have 2 ADV-Micro-Device-DOWN,Andrew-Corp-DOWN,
similar functionality, or
Computer-Assoc-DOWN,Circuit-City-DOWN,
Compaq-DOWN, EMC-Corp-DOWN, Gen-Inst-DOWN,
Motorola-DOWN,Microsoft-DOWN,Scientific-Atl-DOWN
Technology2-DOWN
z Simple segmentation
Fannie-Mae-DOWN,Fed-Home-Loan-DOWN,
group stocks with similar 3 MBNA-Corp-DOWN,Morgan-Stanley-DOWN Financial-DOWN – Dividing students into different registration groups
price fluctuations Baker-Hughes-UP,Dresser-Inds-UP,Halliburton-HLD-UP, alphabetically, by last name
4 Louisiana-Land-UP,Phillips-Petro-UP,Unocal-UP,
Schlumberger-UP
Oil-UP
z Results of a query
z Summarization
– Groupings are a result of an external specification
– Reduce the size of large
data sets
z Graph partitioning
Clustering precipitation – Some mutual relevance and synergy, but areas are not
in Australia identical
Notion of a Cluster can be Ambiguous Types of Clusterings
z A clustering is a set of clusters
z Important distinction between hierarchical and

partitional sets of clusters
How many clusters? Six Clusters
z Partitional Clustering
– A division data objects into non-overlapping subsets (clusters)
such that each data object is in exactly one subset
z Hierarchical clustering
– A set of nested clusters organized as a hierarchical tree
Two Clusters Four Clusters
Partitional Clustering Hierarchical Clustering
p1
p3 p4
p2
p1 p2 p3 p4
Traditional Hierarchical Clustering Traditional Dendrogram
p1
p3 p4
p2
p1 p2 p3 p4
Original Points A Partitional Clustering
Non-traditional Hierarchical Clustering Non-traditional Dendrogram
Other Distinctions Between Sets of Clusters Types of Clusters
z Exclusive versus non-exclusive z Well-separated clusters

– In non-exclusive clusterings, points may belong to multiple
clusters.
– Can represent multiple classes or ‘border’ points z Center-based clusters
z Fuzzy versus non-fuzzy
– In fuzzy clustering, a point belongs to every cluster with some z Contiguous clusters
weight between 0 and 1
– Weights must sum to 1
– Probabilistic clustering has similar characteristics z Density-based clusters
z Partial versus complete
– In some cases, we only want to cluster some of the data z Property or Conceptual
z Heterogeneous versus homogeneous
– Cluster of widely different sizes, shapes, and densities z Described by an Objective Function
Types of Clusters: Well-Separated Types of Clusters: Center-Based
z Well-Separated Clusters: z Center-based

– A cluster is a set of points such that any point in a cluster is – A cluster is a set of objects such that an object in a cluster is
closer (or more similar) to every other point in the cluster than closer (more similar) to the “center” of a cluster, than to the
to any point not in the cluster. center of any other cluster
– The center of a cluster is often a centroid, the average of all
the points in the cluster, or a medoid, the most “representative”
point of a cluster
3 well-separated clusters 4 center-based clusters
Types of Clusters: Contiguity-Based Types of Clusters: Density-Based
z Contiguous Cluster (Nearest neighbor or z Density-based

Transitive) – A cluster is a dense region of points, which is separated by
– A cluster is a set of points such that a point in a cluster is low-density regions, from other regions of high density.
closer (or more similar) to one or more other points in the – Used when the clusters are irregular or intertwined, and when
cluster than to any point not in the cluster. noise and outliers are present.
8 contiguous clusters 6 density-based clusters
Types of Clusters: Conceptual Clusters Types of Clusters: Objective Function
z Shared Property or Conceptual Clusters z Clusters Defined by an Objective Function

– Finds clusters that share some common property or represent – Finds clusters that minimize or maximize an objective function.
a particular concept.
– Enumerate all possible ways of dividing the points into clusters and
. evaluate the `goodness' of each potential set of clusters by using
the given objective function. (NP Hard)
– Can have global or local objectives.
Hierarchical clustering algorithms typically have local objectives
Partitional algorithms typically have global objectives
– A variation of the global objective function approach is to fit the
data to a parameterized model.
Parameters for the model are determined from the data.
Mixture models assume that the data is a ‘mixture' of a number of
statistical distributions.
2 Overlapping Circles
Types of Clusters: Objective Function … Characteristics of the Input Data Are Important
z Map the clustering problem to a different domain z Type of proximity or density measure
– This is a derived measure, but central to clustering
and solve a related problem in that domain
z Sparseness
– Proximity matrix defines a weighted graph, where the – Dictates type of similarity
nodes are the points being clustered, and the – Adds to efficiency
weighted edges represent the proximities between z Attribute type
points – Dictates type of similarity
z Type of Data
– Clustering is equivalent to breaking the graph into – Dictates type of similarity
connected components, one for each cluster. – Other characteristics, e.g., autocorrelation
z Dimensionality
– Want to minimize the edge weight between clusters z Noise and Outliers
and maximize the edge weight within clusters z Type of Distribution
Clustering Algorithms K-means Clustering
z K-means and its variants z Partitional clustering approach

z Each cluster is associated with a centroid (center point)
z Each point is assigned to the cluster with the closest
z Hierarchical clustering centroid
z Number of clusters, K, must be specified
z Density-based clustering z The basic algorithm is very simple
K-means Clustering – Details Two different K-means Clusterings

z Initial centroids are often chosen randomly. 3
2.5
– Clusters produced vary from one run to another. 2

Original Points
z The centroid is (typically) the mean of the points in the 1.5
y
cluster. 1
0.5
z ‘Closeness’ is measured by Euclidean distance, cosine 0
similarity, correlation, etc. -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

x
z K-means will converge for common similarity measures

mentioned above. 3 3
z Most of the convergence happens in the first few 2.5 2.5
iterations. 2 2
1.5 1.5
– Often the stopping condition is changed to ‘Until relatively few

y
y
1 1
points change clusters’ 0.5 0.5
z Complexity is O( n * K * I * d ) 0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

– n = number of points, K = number of clusters, x x
I = number of iterations, d = number of attributes Optimal Clustering Sub-optimal Clustering
Importance of Choosing Initial Centroids Importance of Choosing Initial Centroids

Iteration 1 Iteration 2 Iteration 3
3 3 3
Iteration 6
1
2
3
4
5 2.5 2.5 2.5
3
2 2 2
1.5 1.5 1.5

2.5
y
1 1 1
2 0.5 0.5 0.5
0 0 0
1.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x x x
y

3 3 3
0.5
2.5 2.5 2.5
0 2 2 2
1.5 1.5 1.5

y
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 1 1 1
x 0.5 0.5 0.5
0 0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x x x
Evaluating K-means Clusters Importance of Choosing Initial Centroids …
z Most common measure is Sum of Squared Error (SSE) Iteration 5

1
2
3
4
– For each point, the error is the distance to the nearest cluster 3
– To get SSE, we square these errors and sum them. 2.5
K
SSE = ∑ ∑ dist 2 ( mi , x )
2
i =1 x∈Ci 1.5
y
– x is a data point in cluster Ci and mi is the representative point for 1
cluster Ci
can show that mi corresponds to the center (mean) of the cluster 0.5
– Given two clusters, we can choose the one with the smallest 0
error
– One easy way to reduce SSE is to increase K, the number of -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
clusters x
A good clustering with smaller K can have a lower SSE than a poor
clustering with higher K
Importance of Choosing Initial Centroids … Problems with Selecting Initial Points
3
Iteration 1
3
Iteration 2
z If there are K ‘real’ clusters then the chance of selecting
2.5 2.5 one centroid from each cluster is small.
2 2
1.5 1.5
– Chance is relatively small when K is large
y
1 1 – If clusters are the same size, n, then

0.5 0.5
0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

x x

3 3 3
2.5 2.5 2.5

– For example, if K = 10, then probability = 10!/1010 = 0.00036
2 2 2
1.5 1.5 1.5

– Sometimes the initial centroids will readjust themselves in
‘right’ way, and sometimes they don’t
y
1 1 1
0.5 0.5 0.5

– Consider an example of five pairs of clusters
0 0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x x x
10 Clusters Example 10 Clusters Example

Iteration 4
1
2
3 8
Iteration 1
8
Iteration 2
8
6 6
4 4
6
2 2
y
0 0
4
-2 -2
-4 -4
2
-6 -6
0 5 10 15 20 0 5 10 15 20
y
0 x x
Iteration 3 Iteration 4
8 8
-2 6 6
4 4
-4 2 2
y
0 0
-6 -2 -2
-4 -4
-6 -6
0 5 10 15 20
x 0 5 10
x
15 20 0 5 10
x
15 20
Starting with two initial centroids in one cluster of each pair of clusters Starting with two initial centroids in one cluster of each pair of clusters
10 Clusters Example 10 Clusters Example
Iteration 1 Iteration 2
Iteration 4
1
2
3 8 8
8 6 6
4 4
6 2 2
y
0 0
4 -2 -2
-4 -4
2 -6 -6
0 5 10 15 20 0 5 10 15 20
y
x
Iteration 3 x
Iteration 4
0 8 8
6 6
-2 4 4
2 2
-4
y
0 0
-2 -2
-6 -4 -4
-6 -6
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
x x
x
Starting with some pairs of clusters having three initial centroids, while other have only one. Starting with some pairs of clusters having three initial centroids, while other have only one.
Solutions to Initial Centroids Problem Handling Empty Clusters
z Multiple runs z Basic K-means algorithm can yield empty

– Helps, but probability is not on your side clusters
z Sample and use hierarchical clustering to
determine initial centroids z Several strategies
z Select more than k initial centroids and then – Choose the point that contributes most to SSE
select among these initial centroids – Choose a point from the cluster with the highest SSE
– Select most widely separated – If there are several empty clusters, the above can be
z Postprocessing repeated several times.
z Bisecting K-means
– Not as susceptible to initialization issues
Updating Centers Incrementally Pre-processing and Post-processing
z In the basic K-means algorithm, centroids are z Pre-processing

updated after all points are assigned to a centroid – Normalize the data
– Eliminate outliers
z An alternative is to update the centroids after
each assignment (incremental approach) z Post-processing
– Eliminate small clusters that may represent outliers
– Each assignment updates zero or two centroids
– Split ‘loose’ clusters, i.e., clusters with relatively high
– More expensive
SSE
– Introduces an order dependency
– Merge clusters that are ‘close’ and that have relatively
– Never get an empty cluster low SSE
– Can use “weights” to change the impact – Can use these steps during the clustering process
ISODATA
Bisecting K-means Bisecting K-means Example
z Bisecting K-means algorithm

– Variant of K-means that can produce a partitional or a
hierarchical clustering
Limitations of K-means Limitations of K-means: Differing Sizes
z K-means has problems when clusters are of

differing
– Sizes
– Densities
– Non-globular shapes
z K-means has problems when the data contains

outliers.
Original Points K-means (3 Clusters)
Limitations of K-means: Differing Density Limitations of K-means: Non-globular Shapes
Original Points K-means (3 Clusters) Original Points K-means (2 Clusters)
Overcoming K-means Limitations Overcoming K-means Limitations
Original Points K-means Clusters Original Points K-means Clusters
One solution is to use many clusters.

Find parts of clusters, but need to put together.
Overcoming K-means Limitations Hierarchical Clustering
z Produces a set of nested clusters organized as a

hierarchical tree
z Can be visualized as a dendrogram
– A tree like diagram that records the sequences of
merges or splits
6 5
0.2
4
3 4
0.15 2
5
2
0.1
Original Points K-means Clusters
1
0.05
3 1
0
1 3 2 5 4 6
Strengths of Hierarchical Clustering Hierarchical Clustering
z Do not have to assume any particular number of z Two main types of hierarchical clustering
clusters – Agglomerative:
– Any desired number of clusters can be obtained by Start with the points as individual clusters
‘cutting’ the dendogram at the proper level At each step, merge the closest pair of clusters until only one cluster
(or k clusters) left
z They may correspond to meaningful taxonomies – Divisive:

– Example in biological sciences (e.g., animal kingdom, Start with one, all-inclusive cluster
phylogeny reconstruction, …) At each step, split a cluster until each cluster contains a point (or
there are k clusters)
z Traditional hierarchical algorithms use a similarity or

distance matrix
– Merge or split one cluster at a time
Agglomerative Clustering Algorithm Starting Situation
z More popular hierarchical clustering technique z Start with clusters of individual points and a
z Basic algorithm is straightforward proximity matrix p1 p2 p3 p4 p5 ...
1. Compute the proximity matrix p1
2. Let each data point be a cluster p2
3. Repeat p3
4. Merge the two closest clusters p4

p5
5. Update the proximity matrix .
6. Until only a single cluster remains .
. Proximity Matrix
z Key operation is the computation of the proximity of
two clusters
– Different approaches to defining the distance between
clusters distinguish the different algorithms
Intermediate Situation Intermediate Situation
z After some merging steps, we have some clusters z We want to merge the two closest clusters (C2 and C5) and
C1 C2 C3 C4 C5 update the proximity matrix. C1 C2 C3 C4 C5
C1 C1
C2 C2
C3 C3
C3 C3
C4 C4
C4 C4
C5 C5
Proximity Matrix Proximity Matrix

C1 C1
C2 C5 C2 C5
After Merging How to Define Inter-Cluster Similarity

p1 p2 p3 p4 p5 ...
z The question is “How do we update the proximity matrix?”
p1
C2 Similarity?
U
C1 C5 C3 C4 p2
C1 ? p3
C2 U C5 ? ? ? ? p4
C3
C3 ? p5
C4 z MIN
? .
C4 z MAX
.
Proximity Matrix z Group Average
C1 .
Proximity Matrix
z Distance Between Centroids
z Other methods driven by an objective
C2 U C5
function
– Ward’s Method uses squared error
How to Define Inter-Cluster Similarity How to Define Inter-Cluster Similarity
p1 p2 p3 p4 p5 ... p1 p2 p3 p4 p5 ...
p1 p1
p2 p2
p3 p3
p4 p4
p5 p5
z MIN z MIN
. .
z MAX z MAX
. .
z Group Average .
z Group Average .
z Distance Between Centroids z Distance Between Centroids
z Other methods driven by an objective z Other methods driven by an objective
function function
– Ward’s Method uses squared error – Ward’s Method uses squared error
How to Define Inter-Cluster Similarity How to Define Inter-Cluster Similarity

p1 p2 p3 p4 p5 ... p1 p2 p3 p4 p5 ...
p1 p1
p2 × × p2
p3 p3
p4 p4
p5 p5
z MIN z MIN
. .
z MAX z MAX
. .
z Group Average .
z Group Average .
z Distance Between Centroids z Distance Between Centroids
z Other methods driven by an objective z Other methods driven by an objective
function function
– Ward’s Method uses squared error – Ward’s Method uses squared error
Cluster Similarity: MIN or Single Link Hierarchical Clustering: MIN
z Similarity of two clusters is based on the two

5
most similar (closest) points in the different 1
clusters 3
– Determined by one pair of points, i.e., by one link in 5 0.2
2 1
the proximity graph. 0.15
2 3 6 0.1
I1 I2 I3 I4 I5 0.05
I1 1.00 0.90 0.10 0.65 0.20 4

I2 0.90 1.00 0.70 0.60 0.50 4 0
3 6 2 5 4 1
I3 0.10 0.70 1.00 0.40 0.30

I4 0.65 0.60 0.40 1.00 0.80
I5 0.20 0.50 0.30 0.80 1.00 1 2 3 4 5 Nested Clusters Dendrogram
Strength of MIN Limitations of MIN
Original Points Two Clusters
• Can handle non-elliptical shapes • Sensitive to noise and outliers
Cluster Similarity: MAX or Complete Linkage Hierarchical Clustering: MAX
z Similarity of two clusters is based on the two least

similar (most distant) points in the different 4 1
clusters 2 5 0.4
0.35
– Determined by all pairs of points in the two clusters 5 0.3
2
0.25
3 6 0.2
I1 I2 I3 I4 I5 3
1
0.15
0.1
I1 1.00 0.90 0.10 0.65 0.20 4 0.05
I2 0.90 1.00 0.70 0.60 0.50 0

3 6 4 1 2 5
I3 0.10 0.70 1.00 0.40 0.30

I4 0.65 0.60 0.40 1.00 0.80 Nested Clusters Dendrogram
I5 0.20 0.50 0.30 0.80 1.00 1 2 3 4 5
Strength of MAX Limitations of MAX

• Less susceptible to noise and outliers •Tends to break large clusters

•Biased towards globular clusters
Cluster Similarity: Group Average Hierarchical Clustering: Group Average
z Proximity of two clusters is the average of pairwise proximity

between points in the two clusters.
∑ proximity(p , p )
pi∈Clusteri
i j
5 4 1
p j∈Clusterj
proximity(Clusteri , Clusterj ) = 2 0.25
|Clusteri |∗|Clusterj |
5 0.2
2
z Need to use average connectivity for scalability since total 0.15
proximity favors large clusters 3 6 0.1
1
I1 I2 I3 I4 I5 0.05
4
I1 1.00 0.90 0.10 0.65 0.20 3
0
3 6 4 1 2 5
I2 0.90 1.00 0.70 0.60 0.50

I3 0.10 0.70 1.00 0.40 0.30
I4 0.65 0.60 0.40 1.00 0.80 Nested Clusters Dendrogram
I5 0.20 0.50 0.30 0.80 1.00 1 2 3 4 5
Hierarchical Clustering: Group Average Cluster Similarity: Ward’s Method
z Compromise between Single and Complete z Similarity of two clusters is based on the increase
Link in squared error when two clusters are merged
– Similar to group average if distance between points is
distance squared
z Strengths
– Less susceptible to noise and outliers z Less susceptible to noise and outliers
z Biased towards globular clusters

z Limitations
– Biased towards globular clusters z Hierarchical analogue of K-means
– Can be used to initialize K-means
Hierarchical Clustering: Comparison Hierarchical Clustering: Time and Space requirements
1
5
4 1
z O(N2) space since it uses the proximity matrix.
3
5
2 5 – N is the number of points.
2 1 5
MIN MAX 2
2 3 6 3 6
3
4
1 z O(N3) time in many cases
4
4
– There are N steps and at each step the size, N2,
proximity matrix must be updated and searched
5
1 5 4 1 – Complexity can be reduced to O(N2 log(N) ) time for
2 2 some approaches
5 Ward’s Method 5
2 2
3 6 Group Average 3 6
3
4 1 1
4 4
3
Hierarchical Clustering: Problems and Limitations MST: Divisive Hierarchical Clustering
z Once a decision is made to combine two clusters, z Build MST (Minimum Spanning Tree)
it cannot be undone – Start with a tree that consists of any point
– In successive steps, look for the closest pair of points (p, q) such
that one point (p) is in the current tree but the other (q) is not
z No objective function is directly minimized – Add q to the tree and put an edge between p and q
z Different schemes have problems with one or

more of the following:
– Sensitivity to noise and outliers
– Difficulty handling different sized clusters and convex
shapes
– Breaking large clusters
MST: Divisive Hierarchical Clustering DBSCAN
z Use MST for constructing hierarchy of clusters z DBSCAN is a density-based algorithm.

– Density = number of points within a specified radius (Eps)
– A point is a core point if it has more than a specified number

of points (MinPts) within Eps
These are points that are at the interior of a cluster
– A border point has fewer than MinPts within Eps, but is in

the neighborhood of a core point
– A noise point is any point that is not a core point or a border

point.
DBSCAN: Core, Border, and Noise Points DBSCAN Algorithm
z Eliminate noise points

z Perform clustering on the remaining points
DBSCAN: Core, Border and Noise Points When DBSCAN Works Well
Original Points Clusters
Original Points Point types: core,

border and noise • Resistant to Noise
• Can handle clusters of different shapes and sizes
Eps = 10, MinPts = 4
When DBSCAN Does NOT Work Well DBSCAN: Determining EPS and MinPts
z Idea is that for points in a cluster, their kth nearest

neighbors are at roughly the same distance
z Noise points have the kth nearest neighbor at farther
distance
z So, plot sorted distance of every point to its kth
nearest neighbor
(MinPts=4, Eps=9.75).
Original Points
• Varying densities
• High-dimensional data
(MinPts=4, Eps=9.92)
Cluster Validity Clusters found in Random Data

1 1
z For supervised classification we have a variety of 0.9 0.9
measures to evaluate how good our model is 0.8 0.8
0.7 0.7
– Accuracy, precision, recall Random 0.6 0.6 DBSCAN
Points 0.5 0.5
y
0.4 0.4
z For cluster analysis, the analogous question is how to 0.3 0.3
evaluate the “goodness” of the resulting clusters? 0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
z But “clusters are in the eye of the beholder”! 1

x
1
x
0.9 0.9
0.8
K-means 0.8
Complete
z Then why do we want to evaluate them? 0.7 0.7
Link
– To avoid finding patterns in noise 0.6
0.5
0.6
0.5
y
– To compare clustering algorithms 0.4 0.4
– To compare two sets of clusters 0.3 0.3
0.2 0.2
– To compare two clusters 0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
Different Aspects of Cluster Validation Measures of Cluster Validity
1. Determining the clustering tendency of a set of data, i.e., z Numerical measures that are applied to judge various aspects
distinguishing whether non-random structure actually exists in the of cluster validity, are classified into the following three types.
data. – External Index: Used to measure the extent to which cluster labels
2. Comparing the results of a cluster analysis to externally known match externally supplied class labels.
results, e.g., to externally given class labels. Entropy
3. Evaluating how well the results of a cluster analysis fit the data – Internal Index: Used to measure the goodness of a clustering
without reference to external information. structure without respect to external information.
Sum of Squared Error (SSE)
- Use only the data
– Relative Index: Used to compare two different clusterings or
4. Comparing the results of two different sets of cluster analyses to
clusters.
determine which is better.
Often an external or internal index is used for this function, e.g., SSE or
5. Determining the ‘correct’ number of clusters. entropy
z Sometimes these are referred to as criteria instead of indices
For 2, 3, and 4, we can further distinguish whether we want to – However, sometimes criterion is the general strategy and index is the
numerical measure that implements the criterion.
evaluate the entire clustering or just individual clusters.
Measuring Cluster Validity Via Correlation Measuring Cluster Validity Via Correlation
z Two matrices
– Proximity Matrix
z Correlation of incidence and proximity matrices
– “Incidence” Matrix for the K-means clusterings of the following two
One row and one column for each data point data sets.
An entry is 1 if the associated pair of points belong to the same cluster
1 1
An entry is 0 if the associated pair of points belongs to different clusters
0.9 0.9
z Compute the correlation between the two matrices 0.8
0.7
0.8
0.7
– Since the matrices are symmetric, only the correlation between 0.6 0.6
n(n-1) / 2 entries needs to be calculated. 0.5 0.5

y
y
0.4 0.4
z High correlation indicates that points that belong to the 0.3 0.3
0.2 0.2
same cluster are close to each other. 0.1 0.1
0 0
z Not a good measure for some density or contiguity based 0 0.2 0.4
x
0.6 0.8 1 0 0.2 0.4
x
0.6 0.8 1
clusters.
Corr = -0.9235 Corr = -0.5810
Using Similarity Matrix for Cluster Validation Using Similarity Matrix for Cluster Validation
z Order the similarity matrix with respect to cluster

z Clusters in random data are not so crisp
labels and inspect visually.
1 1
1 10
1 0.9 0.9
10 0.9 20 0.8 0.8
0.9
20 0.8 30 0.7 0.7
0.8
40 0.6 0.6
30 0.7
Points
0.7
50 0.5 0.5
y
40 0.6
0.6
Points
60 0.4 0.4
50 0.5
0.5
y
70 0.3 0.3
60 0.4
0.4 80 0.2 0.2
70 0.3 90 0.1 0.1

0.3
80 0.2 100 0 0
0.2 20 40 60 80 100 Similarity 0 0.2 0.4 0.6 0.8 1
90 0.1 Points x
0.1
100 0
0 20 40 60 80 100 Similarity
0 0.2 0.4 0.6 0.8 1
Points
x
DBSCAN
Using Similarity Matrix for Cluster Validation Using Similarity Matrix for Cluster Validation
z Clusters in random data are not so crisp z Clusters in random data are not so crisp
1 1
1 1
10 0.9 0.9
10 0.9 0.9
20 0.8 0.8
20 0.8 0.8
30 0.7 0.7
30 0.7 0.7
40 0.6 0.6
40 0.6 0.6
Points
Points
50 0.5
y
0.5
50 0.5 0.5
y
60 0.4 0.4
60 0.4 0.4
70 0.3 0.3
70 0.3 0.3
80 0.2 0.2
80 0.2 0.2
90 0.1 0.1
90 0.1 0.1
100 0 0
20 40 60 80 100 Similarity 0 0.2 0.4 0.6 0.8 1 100 0 0
20 40 60 80 100 Similarity 0 0.2 0.4 0.6 0.8 1
Points x
Points x
K-means Complete Link
Using Similarity Matrix for Cluster Validation Internal Measures: SSE

z Clusters in more complicated figures aren’t well separated
1
z Internal Index: Used to measure the goodness of a clustering
0.9 structure without respect to external information
1 500
2 6
0.8
– SSE
0.7
1000
4
3 0.6 z SSE is good for comparing two clusterings or two clusters
1500 0.5
0.4
(average SSE).
2000
0.3 z Can also be used to estimate the number of clusters
5 10
0.2
2500
6 9
0.1
7 8
3000 0 4
500 1000 1500 2000 2500 3000 7
2 6
SSE
5
DBSCAN 0
4
-2 3
2
-4
1
-6 0
2 5 10 15 20 25 30
5 10 15
K
Internal Measures: SSE Framework for Cluster Validity
z Need a framework to interpret any measure.

z SSE curve for a more complicated data set – For example, if our measure of evaluation has the value, 10, is that
good, fair, or poor?
z Statistics provide a framework for cluster validity
– The more “atypical” a clustering result is, the more likely it represents
1
2 6 valid structure in the data
3 – Can compare the values of an index that result from random data or
4
clusterings to those of a clustering result.
If the value of the index is unlikely, then the cluster results are valid
5 – These approaches are more complicated and harder to understand.
7
z For comparing the results of two different sets of cluster
analyses, a framework is less necessary.
– However, there is the question of whether the difference between two
SSE of clusters found using K-means index values is significant
Statistical Framework for SSE Statistical Framework for Correlation
z Example z Correlation of incidence and proximity matrices for the

– Compare SSE of 0.005 against three clusters in random data K-means clusterings of the following two data sets.
– Histogram shows SSE of three clusters in 500 sets of random data
points of size 100 distributed over the range 0.2 – 0.8 for x and y
values 1
0.9
1
0.9
0.8 0.8
1 0.7 0.7
50
0.6 0.6
0.9
45
0.5 0.5
y
0.8 0.4
40 0.4
0.7 0.3 0.3

35
0.2 0.2
0.6
30 0.1 0.1
Count
0.5
y
0 0
25 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x
0.4
20
0.3
15
0.2
10 Corr = -0.9235 Corr = -0.5810
0.1
5
0
0 0.2 0.4 0.6 0.8 1 0
0.016 0.018 0.02 0.022 0.024 0.026 0.028 0.03 0.032 0.034
x SSE
Internal Measures: Cohesion and Separation Internal Measures: Cohesion and Separation
z Cluster Cohesion: Measures how closely related z Example: SSE

are objects in a cluster – BSS + WSS = constant
– Example: SSE m
z Cluster Separation: Measure how distinct or well- × × ×
separated a cluster is from other clusters 1 m1 2 3 4 m2 5
z Example: Squared Error

K=1 cluster: WSS= (1 − 3) 2 + ( 2 − 3) 2 + ( 4 − 3) 2 + (5 − 3) 2 = 10
– Cohesion is measured by the within cluster sum of squares (SSE)
WSS = ∑ ∑ ( x − mi ) 2 BSS= 4 × (3 − 3) 2 = 0
i x∈C i Total = 10 + 0 = 10
– Separation is measured by the between cluster sum of squares
K=2 clusters: WSS= (1 − 1.5) 2 + ( 2 − 1.5) 2 + ( 4 − 4.5) 2 + (5 − 4.5) 2 = 1
BSS = ∑ Ci ( m − mi ) 2
BSS= 2 × (3 − 1.5) 2 + 2 × ( 4.5 − 3) 2 = 9
i
– Where |Ci| is the size of cluster i Total = 1 + 9 = 10
Internal Measures: Cohesion and Separation Internal Measures: Silhouette Coefficient
z A proximity graph based approach can also be used for z Silhouette Coefficient combine ideas of both cohesion and separation,
cohesion and separation. but for individual points, as well as clusters and clusterings
– Cluster cohesion is the sum of the weight of all links within a cluster. z For an individual point, i
– Cluster separation is the sum of the weights between nodes in the cluster – Calculate a = average distance of i to the points in its cluster
and nodes outside the cluster. – Calculate b = min (average distance of i to points in another cluster)
– The silhouette coefficient for a point is then given by
s = 1 – a/b if a < b, (or s = b/a - 1 if a ≥ b, not the usual case)
b
– Typically between 0 and 1. a
– The closer to 1 the better.
cohesion separation z Can calculate the Average Silhouette width for a cluster or a
clustering
External Measures of Cluster Validity: Entropy and Purity Final Comment on Cluster Validity
“The validation of clustering structures is the most

difficult and frustrating part of cluster analysis.
Without a strong effort in this direction, cluster
analysis will remain a black art accessible only to
those true believers who have experience and
great courage.”
Algorithms for Clustering Data, Jain and Dubes

A Clusterng K-Means VCLASS 4

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Clusterng K-Means VCLASS 4

Uploaded by

Copyright:

Available Formats

Data Mining What is Cluster Analysis?

Cluster Analysis: Basic Concepts

Applications of Cluster Analysis What is not Cluster Analysis?

and proteins that have 2 ADV-Micro-Device-DOWN,Andrew-Corp-DOWN,

Notion of a Cluster can be Ambiguous Types of Clusterings

z A clustering is a set of clusters

z Important distinction between hierarchical and

Other Distinctions Between Sets of Clusters Types of Clusters

z Exclusive versus non-exclusive z Well-separated clusters

Types of Clusters: Well-Separated Types of Clusters: Center-Based

z Well-Separated Clusters: z Center-based

3 well-separated clusters 4 center-based clusters

z Contiguous Cluster (Nearest neighbor or z Density-based

8 contiguous clusters 6 density-based clusters

Types of Clusters: Conceptual Clusters Types of Clusters: Objective Function

z Shared Property or Conceptual Clusters z Clusters Defined by an Objective Function

z K-means and its variants z Partitional clustering approach

K-means Clustering – Details Two different K-means Clusterings

– Clusters produced vary from one run to another. 2

z ‘Closeness’ is measured by Euclidean distance, cosine 0

similarity, correlation, etc. -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

z K-means will converge for common similarity measures

z Most of the convergence happens in the first few 2.5 2.5

– Often the stopping condition is changed to ‘Until relatively few

points change clusters’ 0.5 0.5

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

I = number of iterations, d = number of attributes Optimal Clustering Sub-optimal Clustering

Importance of Choosing Initial Centroids Importance of Choosing Initial Centroids

1.5 1.5 1.5

2 0.5 0.5 0.5

Iteration 4 Iteration 5 Iteration 6

1.5 1.5 1.5

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 1 1 1

x 0.5 0.5 0.5

z Most common measure is Sum of Squared Error (SSE) Iteration 5

– To get SSE, we square these errors and sum them. 2.5

Importance of Choosing Initial Centroids … Problems with Selecting Initial Points

1 1 – If clusters are the same size, n, then

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Iteration 3 Iteration 4 Iteration 5

2.5 2.5 2.5

1.5 1.5 1.5

0.5 0.5 0.5

10 Clusters Example 10 Clusters Example

Solutions to Initial Centroids Problem Handling Empty Clusters

z Multiple runs z Basic K-means algorithm can yield empty

Updating Centers Incrementally Pre-processing and Post-processing

z In the basic K-means algorithm, centroids are z Pre-processing

z Bisecting K-means algorithm

Limitations of K-means Limitations of K-means: Differing Sizes

z K-means has problems when clusters are of

z K-means has problems when the data contains

Limitations of K-means: Differing Density Limitations of K-means: Non-globular Shapes

Original Points K-means (3 Clusters) Original Points K-means (2 Clusters)

Original Points K-means Clusters Original Points K-means Clusters

One solution is to use many clusters.

Overcoming K-means Limitations Hierarchical Clustering

z Produces a set of nested clusters organized as a

Strengths of Hierarchical Clustering Hierarchical Clustering

z They may correspond to meaningful taxonomies – Divisive:

z Traditional hierarchical algorithms use a similarity or

2. Let each data point be a cluster p2

4. Merge the two closest clusters p4

Intermediate Situation Intermediate Situation