You are on page 1of 10

Collaboration Recommendation on

Academic Social Networks

Giseli Rabello Lopes1 , Mirella M. Moro2 , Leandro Krug Wives1 ,


and José Palazzo Moreira de Oliveira1,
1
Universidade Federal do Rio Grande do Sul - UFRGS
Porto Alegre, Brazil
{grlopes,wives,palazzo}@inf.ufrgs.br
2
Universidade Federal de Minas Gerais - UFMG
Belo Horizonte, Brazil
mirella@dcc.ufmg.br

Abstract. In the academic context, scientific research works are often


performed through collaboration and cooperation between researchers
and research groups. Researchers work in various subjects and in several
research areas. Identifying new partners to execute joint research and
analyzing the level of cooperation of the current partners can be very
complex tasks. Recommendation of new collaborations may be a valuable
tool for reinforcing and discovering such partners. This paper presents
an innovative approach to recommend collaborations on the context of
academic Social Networks. Specifically, we introduce the architecture for
such approach and the metrics involved in recommending collaborations.
We also present an initial case study to validate our approach.

1 Introduction

Nowadays, information can be electronically accessed as soon as they are pub-


lished on the Web. However, problems associated to the information overload
phenomena emerged. The recovery of relevant digital information on the Web is
a complex task and research in the information filtering area, specifically about
recommender systems, is very important.
Recommender systems reduce the problems associated to the information
overload phenomena by minimizing the time spent to access relevant informa-
tion. Recommender systems involve information personalization. Personalization
is related to the ways in which information and services can be tailored to match
the specific needs of a single user or community [1].
However, recommender systems are inserted in a social context since the rec-
ommendations are delivered to a user or a community of users. Perugini et al.
[2] emphasize that recommendation has an inherently social element and is ul-
timately intended to connect people. In this perspective, the social interactions
and its relational aspects must be considered.

This work was partially supported by CNPq, Brazil.

J. Trujillo et al. (Eds.): ER 2010 Workshops, LNCS 6413, pp. 190–199, 2010.
c Springer-Verlag Berlin Heidelberg 2010
Collaboration Recommendation on Academic Social Networks 191

In this context, the research area of Social Networks has emerged. Specifically,
the Social Network Analysis (SNA) is based on the assumption that the relation-
ship’s importance between interaction units is a central point to the evaluation
and the analysis of social interactions. The increasing interest in researching on
Social Networks was encouraged by the popularization of online Social Networks,
which are very interesting Web applications. Nowadays, this type of network is
commonly used, and each network connects millions of users. Examples of online
Social Networks include MySpace, Facebook, Hi5, Orkut, among others. Another
example of the Social Network’s concept application is a co-authorship Social
Network representing a scientific collaboration network.
In this “social growth of the web”, the number of studies involving Social Net-
work Analysis and the number of applications aiming to improve recommender
system have increased. This fact shall lead to advances in both areas and en-
courages the improvement or the development of new approaches to analyze
connections and recommend new ones. In this context, our paper presents an
innovative approach to recommend collaborations on the context of academic
Social Networks. The contributions of this paper can be summarized as follows:
we introduce the architecture for such approach and the metrics involved in rec-
ommending collaborations (Sect. 2) and we present the results of an initial case
study to validate our approach (Sect. 3).

2 Collaborations Recommendation
In this section, we give an overview of our approach to perform the recommenda-
tion of collaborations on the context of academic Social Networks, as illustrated
in Fig. 1. The first step consists on selecting the target user of the recommenda-
tion and the collaborative network to be used (i.e., a group of individuals from a
digital community such as DBLP1 ) (1). Then, the Social Network formed by se-
lected individuals (or a subgroup of them) is constructed and the weights of the
relationships among them are attributed (2). These weights are calculated based
on the information about the publications of the authors obtained from exist-
ing Digital Library (3). These weights indicate the level of “global cooperation”
between pairs of the researchers involved.
Indeed, the researchers’ profiles (profiles of the authors of the collaborative
network) to be used by the recommender system are constructed based on: the
information available about these authors in some digital library (4) and the
classification of the authors’ publications made with the aid of a research area
ontology (taxonomy) (5). Further details about this kind of ontology can be ob-
tained in [3]. Our approach considers the research areas in which the researchers
work to determine the “global correlation” index. The “global cooperation” and
the “global correlation” indexes between pairs of authors are then used (6 and 7)
as the basis to generate recommendations of authors with whom the user could
establish new collaborations. Finally, these recommendations are presented to
the user (8). The approach is detailed in the following subsections.
1
DBLP Computer Science Bibliography, http://www.informatik.uni-trier.de/˜ley/db
192 G.R. Lopes et al.

(1) Select User and (2) Collaborative (6)


Collaborative Network
Network
(3)
Collaborations
Recommendation

Digital
(4) Library

(7)
(5)
Research Areas
Researchers Profile

Research
areas
Ontology

(8)

Fig. 1. Overview of the proposed Collaborative Recommender System

2.1 Global Cooperation


Social Networks are based on the relationship’s importance between interaction
units. The interaction units of Social Networks are known as actors. The deter-
mination of relationships’ weights between actors in a Social Network is a great
challenge. These weights measure the importance of the existing relationships
between actors. In this paper, the Social Network that we analyze is a scientific
collaboration network. We present a metric to determine one type of association
namely Collaboration in Co-authorship, which defines the global cooperation be-
tween authors in a collaboration network. Actually, the information available in
this type of Social Network is about co-authored papers between authors.
Formally, a Social Network SN of a co-author relationship a is a pair: SNa =
(N, E) where N and E are the set of N odes and Edges, respectively. Each edge
e ∈ E is a tuple of the form ai , t, w, aj , where the edge is directed from ai to aj ,
t denotes the type of association between ai and aj , and w denotes the weight
affected to the association. The weight is a numerical value between 0 and 1 and
the Collaboration in Co-authorship weight (wtCa ) is given by the equation 1.

|aj co authorship|
wtCa(ai →aj ) = (1)
|ai author|
where:
– wtCa(ai →aj ) corresponds to the weight of the collaboration based on the co-
author relationship (the weight is different according to the relation direc-
tion, i.e., the weight ai → aj is different from aj → ai );
– |aj co authorship| corresponds to the number of times that author aj has
co-authored a paper with ai ;
– |ai author| corresponds to the total number of papers of the author ai .
In other words, the higher the wtCa weight is, the more relevant is the relationship
between author aj to the author ai . The use of wtCa metric implies that there
Collaboration Recommendation on Academic Social Networks 193

is a graph with 0 or 2 links between two authors. The weights represent the
degree of collaboration in co-authorship between these authors. This metric is
an asymmetric variant of the Jaccard Coefficient and it has already been applied
in the context of Social Networks [4,5].

2.2 Global Correlation


For performing co-authorship recommendation, it is important to determine a
metric that states the global correlation between researchers. This correlation
may consider the different research areas in which the involved researchers work.
This was chosen because is an important facet to be considered in this context.
For this purpose, we take two actions: we construct a working profile to each
researcher by considering the research areas in which the researchers work; and
we assign weights to represent the grades of contributions of the researcher in
the corresponding area. The global correlation is calculated between working
profiles of the researchers.
Equation 2 calculates the weight of each research area that will compose the
working profile of the researcher.

|ai authorresearch area x |


wRa (ai , x) = (2)
|ai author|

where |ai authorresearch area x | corresponds to the number of papers that author
ai published in the research area x, and |ai author| to the total number of pa-
pers by ai ’. Then, each area x has a corresponding weight that indicates the
contributions of the researcher in that area.
The weights indicate the grade of contributions of the researcher in each re-
search area, and they are used in the calculation of the global correlation between
pairs of authors. Vector Space Model (VSM) [6] is used in our work to perform
this computation. The VSM uses an n-dimensional space to represent the terms,
i.e. n corresponds to the number of distinct research areas. Each author profile
is represented by a research area’s vector, and the weights represent the vector’s
coordinates in the corresponding dimension. Based on the VSM, similarity is
calculated between pairs of authors and the index terms correspond to the re-
search areas of the authors. The weight assigned to each research area allows to
distinguish between the areas, and it is calculated according to their importance
to the author considered by using equation (2). These weights vary continuously
between 0 and 1. Values near to 1 correspond to more important research areas
while values near to 0 correspond to less important research areas. The VSM
principle is based on the inverse correlation between the distance (angle) among
term vectors in the space and the similarity between the information (in this case
the authors profile) that they represent. To calculate the similarity score, that
will represent the global correlation between two authors, the cosine is used as in
equation 3. The resultant value indicates the global correlation degree between
the two authors profiles (ai and aj ), where wRa (ai , x) represents the weights of
194 G.R. Lopes et al.

Table 1. Actions to be taken according to global cooperation and global correlation


degrees

Cooperation/Correlation low medium high


low ok recommend recommend
medium alert ok alert
high alert alert ok

the research areas that composed the user’s profile, and n represents the total
number of research areas.
n
k=1 wRa (ai , xk ).wRa (aj , xk )
global correlation(ai , aj ) =  (3)
n 2 n 2
k=1 (wRa (ai , xk )) . k=1 (wRa (aj , xk ))

2.3 Recommendation
Our approach to generate cooperation recommendations considers the analysis
of all indexes previously described. The relationship between the “global co-
operation” and the “global correlation” for each pair of authors establish the
necessity (or not) of having more research interaction between them. For this
kind of analysis, we establish degrees to represent the different ranges of val-
ues that are possible for these weights. The degrees are “high”, “medium” and
“low”. Based on this analysis, recommendation actions can be established to
indicate possible cooperations between pairs of authors. The combination be-
tween degrees and the interpretation of them can be observed in Table 1, where
the action of recommending a collaboration is indicated by the term “recom-
mend”. For instance, pairs of authors with high global correlation between them
but with low global cooperation must be recommended to intensify their coop-
eration. Moreover, pairs of authors with low global correlation between them
and high or medium global cooperation could also be recommended. The term
“ok” indicates the ideal relationship. The term “alert” indicates situations in
which the global cooperation and global correlation degrees are different but do
not characterize a recommendation action. For example, pairs of authors with
high or medium global cooperation that have lesser global correlation degrees
do not need to be recommended to cooperate, since the cooperation between
the authors is already happening in a more intense relation than their profiles
correlation suggests.
In the case study presented in Sect. 3, we presented an example of translation
of the ranges of numeric values for global cooperation and global correlation into
the degrees defined as “low”, “medium” and “high”.

3 Case Study
In order to analyze how the proposed approach performs, we consider a case
study from the InWeb (Brazilian National Institute of Science and Technology
Collaboration Recommendation on Academic Social Networks 195

for the Web)2 . The Institute is formed by 27 researchers and their students. All
researchers are professors in a Brazilian major education institution (UFMG,
UFRGS, UFAM and CEFET-MG) with graduate program in Computer Science.
For validating our metrics, we have implemented a tool to automatically gen-
erate a Social Network. This SN was built using information about authors
provided by the DBLP on January 21, 2009. It is important to notice that this
library is exported as an XML document. Instead of using the whole dataset, we
extracted from the library just the papers written by the considered researchers
and published in conference proceedings or in journals (as elements inproceed-
ings or article). This data gathering process summed up 677,345 authors; 692,431
conference proceedings papers and 432,663 journal articles. Such a subset was
chosen because this information is significantly important for representing the
co-author relationship between authors and, consequently, to determine the re-
search collaborations among them.
The actors (vertices or nodes) of the SN can be chosen at will. In here, they are
a subset of authors with scientific papers indexed by the DBLP. The relational
ties (linkages or edges) between actors are the relationships between pairs of
authors. These social ties represent the co-author relationships. The weights of
the linkages are determined by the equation 1 (presented in the Sect. 2). In that
equation, |ai author| corresponds to the total number of papers of the author ai ,
and it considers all papers to this author ai indexed at DBLP (papers published
in conferences proceedings or in journals), including those papers that are not
co-authored by authors in the selected SN (that can be a sub-network of the
global SN). These weights represent the “global cooperation” index between
researchers. The SN will be represented by a directional graph where the edges
have a direction associated, according to the equation previously defined. The
InWeb SN was automatically constructed using our developed system. The result
is presented in Fig. 2.
The recommendation method was implemented as presented in Sect. 2. The
profiles of the researchers (author profiles of the collaborative network) used
by the recommendation method were built based on information available about
these authors in the DBLP and on the classification of authors publications made
using an ontology of research areas. The ontology (taxonomy) for classifying au-
thors papers was proposed by Loh et al. [3]. This ontology uses a classification of
computer science areas similar to the ACM (Association for Computing Machin-
ery) classification3 and provides associated weights to the keywords according
to their relevance to each research area. In this case study, only the keywords
present on the publication’s title were used. The proposed approach considers
the research areas in which researchers work to determine a “global correlation”
index. The “global cooperation” and “global correlation” indexes between pairs
of authors are used as a base to generate recommendations for authors with
whom they should establish new collaborations, or that they must intensify col-
laborations. For the definition of recommendations, we have used the method

2
InWeb, http://www.inweb.org.br/
3
ACM Computing Classification, http://www.acm.org/about/class/ccs98-html
196 G.R. Lopes et al.

Legend:

UFAM
1.1 Altigran Soares da Silva
1.2 Edleno Silva de Moura
1.3 João M. B. Cavalcanti

UFRGS
2.1 Carlos A. Heuser
2.2 José Palazzo Moreira de Oliveira
2.3 Leandro Krug Wives
2.4 Renata de Matos Galante
2.5 Viviane Moreira Orengo

CEFET-MG
3.1 Cristina D. Murta
3.2 Evandrino G. Barros
3.3 Fabiano C. Botelho

UFMG
4.1 Adriano M. Pereira
4.2 Alberto H. F. Laender
4.3 Arnaldo de Albuquerque Araújo
4.4 Berthier A. Ribeiro-Neto
4.5 Clodoveu A. Davis
4.6 Dorgival Olavo Guedes Neto
4.7 Genaı́na Nunes Rodrigues
4.8 Gisele L. Pappa
4.9 Jussara M. Almeida
4.10 Marcos André Gonalves
4.11 Mirella Moura Moro
4.12 Nivio Ziviani
4.13 Raquel Oliveira Prates
4.14 Renato Ferreira
4.15 Virgı́lio A. F. Almeida
4.16 Wagner Meira Jr.

Fig. 2. Automatic InWeb Social Network

presented in Sect. 2.3. We have used a linear scale to translate the ranges of
numeric values for global cooperation and global correlation into the degrees
defined as “low”, “medium” and “high”. The “low”, “medium” and “high” de-
grees correspond, respectively, to values v in the ranges {v ∈ |0.0 < v ≤ 1/3},
{v ∈ |1/3 < v ≤ 2/3} and {v ∈ |2/3 < v ≤ 1.0}.
After the identification of the resulting degree, it is possible to establish recom-
mendation actions according to Table 1. As shown in that table, the recommen-
dation action will only happen when there is a “low” value of global cooperation
between pairs of authors and a “medium” or “high” value of global correla-
tion. Moreover, when the value of global cooperation is effectively null (zero), a
recommendation to initiate collaboration will be made to the authors. Indeed,
when the value of global cooperation is a “low” but not null, an intensification
of existing collaboration will be recommended.
An example of interface for presenting recommendations for the researcher
“José Palazzo Moreira de Oliveira” (InWeb member), can be seen in Fig. 3. In
the recommendation to “initiate collaboration”, the value of the global correla-
tion index obtained between the recommended research and the target researcher
is shown. When it is recommended to “intensify cooperation”, the ratio of global
cooperation and global correlation is presented. Furthermore, recommendations
are presented in ranked lists: in the “initiate collaboration” list, the researchers
recommended are in descending order of global correlation index, since the higher
the value of global correlation found, the greater the possibility of collaboration
be interesting to the target researcher; and in the “intensify cooperation” list,
the recommended researchers are presented in increasing order of ratio between
Collaboration Recommendation on Academic Social Networks 197

José Palazzo Moreira de Oliveira


Logout
Initiate Cooperation: Intensify Cooperation:

1. Clodoveu A. Davis 0.804 1. Alberto H. F. Laender 0.052


2. Wagner Meira Jr. 0.632 2. João M. B. Cavalcanti 0.065
3. Adriano M. Pereira 0.572 3. Renata de Matos Galante 0.074
4. Berthier A. Ribeiro-Neto 0.564 4. Leandro Krug Wives 0.245
5. Edleno Silva de Moura 0.554
6. Dorgival Olavo Guedes Neto 0.509
7. Virgílio A. F. Almeida 0.502
8. Altigran Soares da Silva 0.488
9. Nivio Ziviani 0.485
10. Raquel Oliveira Prates 0.461

Fig. 3. Example of interface for recommendations presentation

global cooperation and global correlation, so that the recommendations are pre-
sented in order of intensification necessity.

4 Related Works

Traditionally, recommender systems are studied in three different perspectives


according to the methodologies used to perform recommendation: (i) content-
based, which recommends items classified accordingly to the user profile and
early choices considering semantic issues; (ii) collaborative filtering, which deals
with similarities among users interests including the consideration of structural
aspects; and (iii) hybrid approach, which combines the two to take advantage
of their benefits. The related works presented in this section aim to use Social
Networks on the context of Recommendation Systems.
Many current solutions focus only on structural issues of the Social Network
to generate recommendations. Some examples of these collaborative filtering ap-
proaches are presented bellow. Ogata et al. [7] propose PeCo-Mediator-II system
to search cooperators through a chain of personal connections in a SN. The sys-
tem aims to help the user to find cooperators through the chain of connections
between them. Golbeck et al. [8] present a website that integrates Social Net-
works on the Semantic Web context and the concept of trust to generate films
recommendations. The SN indicates the trust rates between users by the path’s
length between them. Aleman-Meza et al. [4] define a solution to the Conflict of
Interest problem (COI) using Social Networks. The objective is to detect rela-
tionships of COI amongst authors of scientific papers and potential reviewers of
these papers based on public sources such as DBLP and FOAF. Quercia et al.
[9] propose a framework, called FriendSensing, to automatically suggest friends
to users of Social Networks constructed based on the physical proximity between
users identified through mobile devices. Karagiannis et al. [10] analyze emails
exchanged by a certain group of people. For this purpose, they construct a So-
cial Network among these people and suggest the recommendation of “friends of
friends”. Chen et al. [11] propose four algorithms to recommend people on the
198 G.R. Lopes et al.

context of Beehive, an IBM’s Social Network. Two of these algorithms, named


FoF and SONAR, use only structural aspects of the SN.
Some approaches consider both the semantic and the structural issues in the
recommendation method. Some examples of these hybrid approaches are pre-
sented bellow. Kautz et al. [12] present the ReferralWeb system to identify ex-
perts in searches by keywords and to generate a path of social relationships
between a user and the recommended expert. The system models and extracts
existing relationships among people of the Computer Science community by min-
ing public data available on Web documents. McDonald [13] details an evaluation
of two different Social Networks that can be used in a system to recommend in-
dividuals for possible collaborations. The system matches individuals looking
for expertise within people that could have this expertise. Zaiane et al. [14] ex-
plore a Social Network coded within the DBLP database by drawing on a new
random walk approach in order to reveal interesting knowledge about the re-
search community and even recommend collaborations. The approach aims at
helping the user on the process of searching for relevant conferences, similar au-
thors and interesting research topics. Weng et al. [15] propose a recommendation
method that uses ontologies and the spreading activation model. This model is
used to search for other influential users on a Social Network. The ontologies
are used to define the users profiles and as base to infer users’ interests. One of
the algorithms proposed by Chen et al. [11] to recommend people on the Bee-
hive, named CplusL, combine two approaches of recommendations: the approach
based on content and the approach that uses the structural aspects of the SN.
In this paper, we propose an approach to consider the working area of the rec-
ommendation target user by trying to match this information to recommended
a researcher with similar profile. Most of the approaches that perform collab-
orations recommendation focus on recommending experts in a certain area or
information [7,12,13]. They do not consider the working area of the target user.
Our approach also aims at obtaining information about the relationships among
users, implicitly, through publications data obtained from digital libraries. More-
over, it includes not only the recommendation of new collaborations, as well as
the recommendation to intensify existing collaborations, which is not considered
by other related work. Many works focus only on structural issues of the SN to
generate recommendations [4,7,8,9,10,11]. Our work considers both the seman-
tic issues involving the relationship between the researchers in research areas
and the structural issues by the analysis of the existent relationships among
researchers.

5 Final Remarks

In this work, we have presented the details of an innovative approach to rec-


ommend scientific collaborations on the context of Social Networks. Moreover,
we have shown the state of the art and discussed related work. The complete
work is under development as a research project of the InWeb. The case study
for the experiments is a collaborative Social Network based on the publications
Collaboration Recommendation on Academic Social Networks 199

of the researchers associated to the InWeb project. In our case study, the Social
Network is constructed based on data available by the DBLP digital library. In
future works, we will provide more experiments to evaluate the approach aiming
to show the viability and applicability of our proposal in a real world application.

References
1. Smeaton, A.F., Callan, J.: Personalisation and recommender systems in digital
libraries. Int. J. on Digital Libraries 5, 299–308 (2005)
2. Perugini, S., Gonçalves, M.A., Fox, E.A.: Recommender systems research: A
connection-centric survey. J. Intell. Inf. Syst. 23, 107–143 (2004)
3. Loh, S., et al.: Constructing domain ontologies for indexing texts and creating
users’ profiles. In: Work. on Ontologies and Metamodeling in Software and Data
Engineering, Brazilian Symp. on Databases, UFSC, Florianópolis, pp. 72–82 (2006)
4. Aleman-Meza, B., et al.: Semantic analytics on social networks: experiences in
addressing the problem of conflict of interest detection. In: Intl. Conf. on World
Wide Web, pp. 407–416. ACM Press, New York (2006)
5. Mika, P.: Social networks and the semantic web. In: IEEE/WIC/ACM Intl. Conf.
on Web Intelligence, pp. 285–291. IEEE Press, New York (2004)
6. Salton, G., Buckley, C.: Term-weighting approaches in automatic text retrieval.
Inf. Process. Manage. 24, 513–523 (1988)
7. Ogata, H., Yano, Y., Furugori, N., Jin, Q.: Computer supported social networking
for augmenting cooperation. Comput. Supp. Coop. Work 10, 189–209 (2001)
8. Golbeck, J., Hendler, J.: Filmtrust: movie recommendations using trust in web-
based social networks. In: Consumer Communications and Networking Conf., pp.
282–286. IEEE Press, New York (2006)
9. Quercia, D., Capra, L.: Friendsensing: recommending friends using mobile phones.
In: ACM Conference on Recommender Systems, pp. 273–276. ACM Press, New
York (2009)
10. Karagiannis, T., Vojnovic, M.: Behavioral profiles for advanced email features. In:
Intl. Conf. on World Wide Web, pp. 711–720. ACM Press, New York (2009)
11. Chen, J., et al.: Make new friends, but keep the old: recommending people on
social networking sites. In: SIGCHI Conference on Human Factors in Computing
Systems, pp. 201–210. ACM Press, New York (2009)
12. Kautz, H., Selman, B., Shah, M.: Referral web: combining social networks and
collaborative filtering. Commun. ACM 40, 63–65 (1997)
13. McDonald, D.W.: Recommending collaboration with social networks: a compara-
tive evaluation. In: SIGCHI Conference on Human Factors in Computing Systems,
pp. 593–600. ACM Press, New York (2003)
14. Zaiane, O.R., Chen, J., Goebel, R.: Dbconnect: mining research community on
dblp data. In: WebKDD and SNA-KDD Work. on Web Mining and Social Network
Analysis, pp. 74–81. ACM Press, New York (2007)
15. Weng, S.S., Chang, H.L.: Using ontology network analysis for research document
recommendation. Expert Syst. Appl. 34, 1857–1869 (2008)

You might also like