Content uploaded by Tom Z. J. Fu
Author content
All content in this area was uploaded by Tom Z. J. Fu on May 06, 2019
Content may be subject to copyright.
Networking, clustering and brokering keywords in the
computer science research
- Analysis of the Evolution using Social Network Analysis
Jihyoun Park, Tom Z. J. Fu, Dah Ming Chiu
Department of Information Engineering
The Chinese University of Hong Kong, Hong Kong
E-mail: {jhpark, zjfu6, dmchiu}@ie.cuhk.edu.hk
Abstract— Keywords of academic papers are jargons shared
within a research domain as well as a summary of the contents.
However, the increasing popularity of the interdisciplinary
research in academia in recent years opened a possibility that the
choice of keywords would be no longer confined by the
traditional domain boundaries. In order to investigate the
evolutionary trends of keyword-sharing groups in academic
publications, we applied three types of Social Network Analysis
methods on the papers in the Computer Science domain collected
from the Libra database. Particularly, papers published in 5
different time periods from 1970 to 2010 were targeted to show
the change of keyword networks over time. The analysis results
confirmed the increasing tendency in the keyword network to be
absorbed by one big agglomerated cluster over time. As a result,
the gatekeeper- or liaison keywords seldom exist in later times
but coordinator keywords became majority.
Keywords-component; social network analysis; academic
keyword analysis; cluster; brokerage
I. INTRODUCTION
The interdisciplinary research gets more popularity and
attention in recent days, because not only the convergence
among different disciplines can provide more innovative tools
to solve a problem by approaching the old problem with brand-
new aspects, but also the global issues that we are now facing
are not confined by a certain domain boundary. Therefore, the
demand for interdisciplinary research is expected to increase
further.
Keywords of an academic paper are the most representative
words to summarize the contents of the paper. They are used to
categorize the paper as well. Therefore, research papers that
belong to the same or related domain area share keywords; and
we can expect that the relationship would be applicable the
other way around, too, if the domains are well classified to be
exclusive and distinctive. However, with the increase of inter-
and multidisciplinary research, we expect that more and more
research papers share keywords beyond the traditional domain
boundaries. Therefore the boundary of traditionally different
research domains would not be as distinguishable as they were.
This paper investigates the keywords network of research
papers in the computer science discipline. We will analyze the
networking characteristics, clustering structure and brokering
nodes between clusters of the network. Those analyses are
conducted for 5 different time periods from 1970 to 2010 with
10 years interval separating them so that the dynamics of the
keywords network over time can be captured. Through the
analysis, we will test how the keyword-sharing groups formed
and changed and which keywords took the role to broker
different keyword-sharing groups for each time period.
Brokerage roles [1, 2] are defined by a vertex’s strategic
location in the network. Contrary to the role of central nodes,
who have power to control information and resource flows in
the network, these network nodes take a role of bridging other
nodes. If we apply this concept to research networks, we can
easily imagine that those broker nodes are the intermediaries
for interdisciplinary research. Especially, representatives,
gatekeepers and liaisons are a type of shared keywords by
different domains. Those keywords with such roles in early
days are likely to become central keywords in later days
because they can take advantage of being related to more
keywords in other domains. However, not all broker keywords
grow up to be central actors. Many keywords in early days
become out-of-date and perish out. Therefore, we are going to
study how they evolved over the course of time. Particularly 10
interesting keywords in 1970 were selected for further
investigation to trace the change of their status and roles from
1970 to 2010.
The remainder of the paper is organized as follows. In
section 2, we will introduce previous literature related to
investigating social networks in scientific research. In section 3,
selected methods of the social network analysis (SNA), which
is the methodology used for our analysis, will be explained.
Section 4 describes the input data of the analysis. After
showing the analysis results in section 5, we conclude this
paper with discussion and future works in Section 6.
II. RELATED WORKS
The past work on academic social networks mainly focused
on the collaboration among countries, organizations, and
individual researchers [3]. They investigate the networking
properties among target entities using various SNA methods in
order to analyze the knowledge flow, the balance and
dependencies among entities, and the trend of change in
research cooperation.
978-1-4673-2430-4/12/$31.00 ©2012 IEEE
110
In academic social network analysis, the academic
influence of a network entity (inside the research communities
represented as a social network) is usually evaluated based on
its co-authorship and citation activities [3-8]. The attention is
on how the research community and academic social networks
evolve over time as a result of authors' collaborations
Represented by co-author and citation links. The common
interests in this field are in researchers' connections or
influence based on their academic achievements reflected by
their networking positions among other researchers.
On the other hand, co-patents [9, 10], websites links [11],
and project partnerships [12, 13] are also widely-used sources
for academic social network research. There are a few works,
which tried to apply SNA methods for analyzing the process of
knowledge creation, exchanges and sharing [14, 15, 16].
However, their emphasis is also mainly on the interactions and
collaborations among individuals or organizations in terms of
the communities in the networks.
Our research is different from previous works in that it uses
the keywords network derived from the papers’ content-based
relationship to present the overall trends of academic research.
III. METHODS
SNA is a methodology to simply represent the relations
between social actors with nodes and edges (for undirected
graphs) or arcs (for directed graphs) using graph theories. In
this way, researchers can utilize various mathematical methods
to investigate the status, power and interactions among social
actors as well as the overall network structure. In this section,
we introduce the selected social network methods applied for
our analysis.
A. General Network Properties
For the 5 derived keyword networks (one for each time
period), we conduct three types of SNA methodologies which
aim at different aspects of the dedicated social network. The
first type is to look into the network attributes [6, 8, 14, 17]
including distance, density, degree, components and centrality
etc. These metrics that reveal some basic and general
topological network characteristics help us to have some
understanding at the coarse level.
B. p-Clique
The second type is a more detailed level analysis which
focuses on the group of vertices of the network, called the p-
clique analysis. Since cliques are suggested to be structured
around some key vertices (e.g. popular or important keyword in
our study), this helps to find out possible highly cohesive
subgroups or clusters. We apply the following definition of p-
clique in our analysis: p-Clique partitions a network to groups
of close neighbors so that vertices inside the group have at least
proportion p (0<p<1) neighbors.
C. Brokerage role
The third type, the Brokerage Role analysis, is at the most
detailed level. Based on the clustering results derived by the p-
clique method, it is more interesting to investigate the roles that
the vertices at different positions of the network could act as
and the corresponding properties they have. The brokerage role
of a vertex is determined by the relative position of the vertex
in a clustered network. Each vertex can have five brokerage
roles according to given partition, including coordinator,
itinerant broker, representative, gatekeeper and liaison. Figure
1 plots a simple illustration of the 5 kinds of brokerage roles.
(In undirected graphs, a representative is equal to a gatekeeper.)
Figure 1. 5 brokerage roles [18]
IV. DATA
In this section, we explain the data source to retrieve
academic publication information in the computer science
discipline and the pre-processing procedure of the data to make
them suitable for our analysis purpose.
A. Libra Database
We used Microsoft Academic Search [19] (also known as
Libra) as the data source of academic social network for
analysis. Currently, Libra provides publication records for
more than 10 disciplines. In this paper, we pick “Computer
Science”, one of the most representatives to study. Libra
provides an API for data download service and several
visualization features for users to take a glance at the trend of
academic research, top authors, and the network of authors and
papers for each research domain.
For each paper record obtained from Libra, we have some
basic information such as paper title, publication year, authors,
conference/journal, keywords, etc. Although we only focus on
one domain, it still results in a huge number of total paper
records for SNA, which makes the analysis process complex
and intractable.
Therefore, we further simplified the analysis by adding two
filtering conditions: a) paper record whose publication year is
either one of the five time periods will be selected: 1970, 1980,
1990, 2000, 2010, where the different time periods could help
to reveal the change and evolution of the keywords network in
this research discipline; b) paper record to be selected should
have at least one keyword information. In this study, we
directly make use of the results of how keywords are extracted
from papers already done by Libra (in the format of paper-
keyword mapping). Investigating on how Libra achieves this
and discussing on the accuracy and efficiency of related
algorithms are out of the scope of this paper. We will consider
in the future works.
Table I shows the number of papers used for the analysis
and that of distinctive keywords counted for all the papers. The
number of papers and keywords increased exponentially during
the time. In the case of paper numbers, it increased about 3
111
times every 10 year from 1970 to 2000 and about 2.5 times for
keywords.
TABLE I. DATA STATISTICS
1970
1980
1990
2000
2010
# of papers
1,986
6,123
19,157
58,128
64,344
# of keywords
1,342
3,319
7,504
17,319
17,448
B. Data transformation
The collected information is in a bipartite mode, i.e. papers
vs. keywords. Since the study focuses on the keyword relations,
we transformed the bipartite graph, or the 2-mode matrix into
the 1-mode (keyword) matrix. For each pair of keywords, there
is an edge connecting them if they belong to the same paper.
V. ANALYSIS RESULTS
We used Pajek [18], which is especially good at handling
large amount of data, to conduct the SNA on the keywords
network of computer science academic papers.
A. Genreal properties
The general properties of the keywords network are shown
in Table II. The major findings are, first, the keywords network
of the computer science discipline shows the typical small
world property. Even though the number of network nodes is
large but average distance from one node to the others in the
network is only few steps. Second, these networks are very
sparse and highly decentralized. The density of the network is
only 0.08%~0.24%. The average degree and network
centralization indicate that the keyword networks are also
decentralized networks. It means that there is no outstanding
node but the power among network nodes is relatively evenly
distributed. Lastly, the number of components, which are self-
connected sub-networks, is not ignorable. However, the
tendency that keywords get connected more and more is
observed over time. After 2000, more than 90% of keywords
connected with each other thus the connectedness index
improved drastically. This trend shows that the entire network
is moving from a highly segmented structure to the mesh
structure, in which keywords link to each other more closely.
TABLE II. GENERAL NETWORK PROPERTIES
1970
1980
1990
2000
2010
# of nodes
(keywords)
1,342
3,319
7,504
17,319
17,448
Avg. distance
7.04
5.28
3.74
2.91
3.12
Density
0.0010
0.0008
0.0012
0.0024
0.0016
# of
Components
728
1,293
1,408
583
1,219
Largest
component
306
(23%)
1,704
(51%)
5,898
(79%)
16,687
(96%)
16,156
(93%)
Connectedness
0.05
0.26
0.62
0.93
0.86
Avg. degree
1.34
2.79
8.76
42.37
28.48
Network
centralization
0.02
0.03
0.05
0.19
0.11
B. Clusters
We then investigated the clustering composition inside the
main component in order to look into how keyword-sharing
groups naturally arose and changed over time. Figures 2 to 6
show the clustering results of keyword-sharing groups for each
time period. The representative keyword of a cluster stands for
the most degree-central keyword. Different size for nodes is
applied to represent the size of a cluster (=number of nodes).
The width of links indicates the frequency of interactions
between nodes of two clusters.
1) 1970: The clustering analysis of 1970’s keywords
detected 28 clusters out of 306 vertices. The two largest
clusters are covariation function (34) and approximate solution
(33) followed by computer simulation (25) and difference
equation (22). The number of clusters, of which members are
between 10 and 19, is 9. The clusters which have less than 10
members are 15. The average cluster size is 11. The
connectivity between clusters is quite active. Particularly, it is
more visible between approximate solution and computer
simulation, between covariance function and spectrum,
between computational complexity and natural language, and
between computational complexity and turing machine.
Figure 2. Keyword-sharing groups in the main component of the keywords
network in 1970
2) 1980: The total number of vertices increased to 1704
but the number of clusters decreased a little (26). The
composition of clusters became unbalanced. The largest
cluster is the one represend by language production, of which
size is 557. It is followed by 5 leading clusters (lower bound,
distirubted system, pattern recognition, spanning tree, and
visual inspection), of which the number of memebers is
between 149 and 231. We can also detect 5 small groups
(queueing network, optimal design, search algorithm,
symbolic computation, and matrix multiplication) and 15 very
small groups of which the number of members is below 10.
Strong connectivity are detected between clusters of pattern
recognition and langauge production, language production and
lower bound, lower bound and distirubted system, distribute
system and visual inspection, and visual inspection and
spanning tree.
112
Figure 3. Keyword-sharing groups in the main component of the keywords
network in 1980
3) 1990: Significant changes in the clustering structure
from the previous times are deteced. The number of vertices
are now 5898, but only 8 clusters were formed. The cluster
represented by indexing terms is a giant cluster that contains
94% of keywords. It means that almost all keywords are now
shared among papers regardless of their domain categories,
which have been traditionaly considered distinguishable.
Figure 4. Keyword-sharing groups in the main component of the keywords
network in 1990
4) 2000: 7 clusters out of 16687 vertices are found. The
largest cluster includes 99.68% keywords. Interaction among
keywords internalized and networking outside clusters is even
sparser than 1990’s.
Figure 5. Keyword-sharing groups in the main component of the keywords
network in 2000
5) 2010: we can observe 5 clusters out of 16156 vetices.
The cluster composition in 2010 is very much similar with
2010. 99.67% of keywords belong to one cluster. In
conclusion, since almost all keywords belong to a single large
cluster, the clustering approach is no longer very effective in
understanding keyword relationships after 1990.
Figure 6. Keyword-sharing groups in the main component of the keywords
network in 2010
C. Brokerage role
Based on the previous analysis of clustering keywords, this
section diagnoses the role of keywords that they took to
promote activities inside a cluster or between clusters (Table
III). The increase tendency of coordinators tells that research
keywords continuously reinforce internal links inside its own
group. From 1990, the majority of keywords took a role of
coordinator. Gatekeepers and liaisons, which connect different
domains together, increased from 1970 to 1980, but suddenly
decreased after 1990. It is because small keyword-sharing
groups converged into one larger keyword-sharing group
through active sharing of keywords in papers of different
domains.
TABLE III. STATISTICS OF BROKERAGE ROLES
1970
1980
1990
2000
2010
Coordinator
17.32%
28.64%
53.59%
71.13%
68.8
%
Itinerant
broker
4.9
%
8.74%
10.09%
0.02%
0.01%
Gatekeeper
21.9%
27.58%
5.07%
0.3%
0.18%
Liaison
14.38%
16.12%
0.54%
0.04%
0.02%
We assume that the dynamics of keywords over time can
follow four paths. First, they are always popular. They are
fundamental concepts and already established in the early time
so until now the centrality and its roles remain important.
Second, they were once popular but as time passes by, the
research focus moved to other domains. Third, at the beginning
they were not very important but strategically positioned to link
multiple domains so later they become much more influential.
This is a possible course when the popularity of
interdisciplinary research at the current time is taken into
consideration. Fourth, they were bridges among different
113
domains in the past but become obsolete because the kind of
interdisciplinary research was not so successful.
We traced the path of 97 keywords from 1970 to 2010 that
had any brokerage role inside a cluster or between clusters in
1970. We observed the status of those keywords in terms of
degree centrality to see how the keywords evolved. The data
were divided into two groups. The first group is coordinators in
1970 and the other is non-coordinator but those who took a
gatekeeper or liaison role. As depicted in Figure 7, the trends of
the two groups are similar but more members of the
gatekeeper/liaison group remained in the core in 2010. 16 of 53
coordinators ranked on top 5% central keywords in 2010 while
27 out of 44 gatekeeper/liaison keywords ranked on top 5%.
These results indicate that the roles of these keywords which
are aggregating information from various domains could make
them competitive to create new corresponding keywords over
time. Similar results can be derived if we repeatedly investigate
on those root keywords.
Figure 7. Status of brokerage keywords of 1970 in 2010
D. Evolution trace of selected keywords
We selected 10 keywords from 1970 to trace the change of
their roles over time in detail. Those 10 keywords were
selected from our observation of change in research trends in
the computer science field as examples to show dynamics of
research interests in this discipline. Table IV describes the roles
of the selected keywords in the brokerage context. C, I, G, and
L represent coordinator, itinerant broker, gatekeeper and liaison
for each. X means that the keyword was not included in the
main component at the time, thus it was excluded from the
cluster and brokerage role analysis. Many of selected keywords
belonged to this category in 1970 because the keyword network
of 1970 was highly segmented. The value in bracket indicates
the degree centrality of the keyword at the time. Programming
language and computational complexity are coordinators as
well as gatekeepers/liaisons in 1970. That means that they were
already in the central position that links keywords inside the
cluster. We can observe that they kept their importance in the
keywords network until 2010. Speech recognition is an
interesting example. It took a role of liaison in 1970 to connect
different domains. As time passes by, it expanded its influence
and become a highly central keyword in later times. The
centrality rank of those keywords that existed in other
components apart from the main component was relatively low
but once they joined the main component, they got recognition
and their rank in degree centrality moved up.
TABLE IV. BROKERAGE ROLES FOR SELECTED KEYWORDS
Keywords
1970
1980
1990
2000
2010
speech
recognition
L
(11.8%)
C,I,G,L
(3.6%)
C
(0.8%)
C
(0.3%)
C
(1.6%)
Programming
Language
C,G,L
(0.4%)
C,I,G,L
(0.1%)
C
(0.1%)
C
(0.2%)
C
(0.4%)
operating
system
(17.4%)
C,I,G,L
(0.4%)
C
(0.2%)
C
(0.2%)
C
(0.9%)
traffic model
X
(15%)
-
(70.1%)
C
(3.7%)
C
(13.2%)
computer
network
X
(19%)
C,G,L
(6.2%)
C
(1.8%)
C
(1.2%)
C
(2.9%)
Computational
Complexity
C,I,G,L
(0.2%)
I,G,L
(5.4%)
C
(0.6%)
C
(0.2%)
C
(0.1%)
dynamic
program
X
(59.2%)
C,G,L
(7.5%)
C
(3.4%)
C
(0.8%)
C
(0.7%)
hash function
X
(65.7%)
(43.6%)
C
(16.3%)
C
(11.3%)
C
(3%)
optical
communication
X
(76.5%)
(25.6%)
C
(53.5%)
C
(15.9%)
C
(20.8%)
stable marriage
X
(76.3%)
-
-
-
C
(33.4%)
a. C: Coordinator, I: Itinerant Broker, G: Gatekeeper, L: Liaison, X: Not defined.
VI. DISCUSSION
In 1970 and 1980, many research communities form small
and independent clusters that use the same jargons. The size of
clustered keyword-sharing groups was relatively evenly
distributed. However, from 1990 to 2010, we observed the
emergence of one agglomerated big community that shared
keywords together. The number of distinct clusters reduced and
relations among clusters become much simpler. These changes
are attributed to strong tendency towards interdisciplinary
research in recent years.
We can also confirm that the time around 1980 was the
most important or fast-changing period to promote inter-
/multidisciplinary research beyond traditional domain
boundaries, since many keywords actively took brokerage roles
not only inside a cluster but between clusters. The development
of gatekeepers and liaisons is particularly important in this time
due to their potential to become innovators to create new inter-
multidisciplinary research by taking advantage of the ability to
link different domains. After 1990, the keywords-network
already formed a big component and has not shown much
dynamics. Therefore, most of keywords become coordinators.
The increasing tendency towards interdisciplinary research
is calling for a review or renewal on the academic domain
classification. With the emergence of new domains and
methodologies in research areas, the classification system has
also been adapted somehow such that some domains were
created and some perished. However, the academic system
seems not flexible enough to follow the ever-changing trend.
By using the predefined hierarchical classification framework,
114
some new research cannot be fully explained by the category
and some cannot even find the right place to fit in. Papers that
were not properly classified might lose the chance to be
properly exposed for search and citation by other researchers
and impact the rapid development of new research ideas.
For future direction, a comparison between our approaches
to derive new classification based on analyzing keyword-
sharing groups, with the traditional classification system is
required. However, for the solution to bridge the gap between
two, an evolutionary approach seems more appropriate than a
revolutionary one. The evolutionary approach can minimize
abrupt changes, and keeps the existing stable domain structure
as a basic framework and adds minor adjustments to it from
time to time; while the revolutionary approach would have
abandoned the entire regime and creates a new domain
structure from scratch.
Through the analysis, we observed that the gatekeeper or
liaison keywords were potential candidates as intermediaries to
connect different domains together and enable the knowledge
flow among them. However, the question is towards
developing a theory of how keywords can dynamically help us
define research domain for classification purposes. In the future,
we will investigate the development of multidisciplinary
research around the gatekeepers or liaisons. However, due to
the huge volume of data, a more systematic approach to deal
with large data must be adopted to test how this type of
keywords develop over time together with other keywords that
have previous contacts with them.
REFERENCES
[1] P. Bonacish, "Power and centrality", American Journal of Socialogy, vol.
92, no. 1, pp. 1170-1182, 1987.
[2] R. S. Burt, Structureal Holes. Cambridge: Harvard University Press,
1992.
[3] W. Glänzel and A. Schubert, "Analyzing scientific networks through co-
authorship", in Handbook of Quantitative Science and Technology
Research, H. F. Moed, W. Glänzel, and U. Schmoch, Eds. Dordrecht,
London: Kluwer Academic Publishers, 2004, pp. 257-276.
[4] C. S. Wagner and L. Leydesdorff, "Mapping the network of global
science: comparing international co-authorships from 1990 to 2000",
International Journal of Technology and Globalisation, vol. 1, no. 2, pp
185-208, 2005.
[5] D. B. Horn, T. A. Finholt, J. P. Birnholtz, D. Motwani, and S. Jayaraman,
"Six degrees of Jonathan Grudin: a social netowrk analysis of the
evolution and impact of CSCW research". Proceedings of the 2004
ACM conference on Computer Supported Cooperative Work (CSCW
'04), New York: ACM, pp. 582-591, 2004.
[6] M. E. J. Newman, "Coauthorship networks and patterns of scientific
collabaration". Proceedings of the National Academy of Sciences, vol.
101, pp. 5200-5205, 2004.
[7] M. Rosvall and C. T. Bergstrom, "Mapping change in large networks",
PLoS ONE, vol. 5, no. 1, 2010.
[8] A. L. Barabasi, H. Jeong, Z. Neda, E. Ravasz, A. Schubert, and T.
Vicsek, "Evolution of the social network of scientific collaborations",
Physica A, vol. 311, pp. 590-614, 2002.
[9] W. Cantner and H. Gra, "The network of innovators in Jena: An
application of Social Network Analysis", Research Policy, vol. 35, no. 4,
pp. 463-480, 2006.
[10] J. Owen-Smith, M. Riccaboni, F. Pammolli, and W. W. Powell, "A
Comparison of US and European University-Industry relations in the
Life Sciences", Management Science, vol. 48, no. 1, pp 24-43, 2002.
[11] J. L. Ortega, I. Aguillo, V. Cothey, and A. Scharnhorst, "Maps of the
academic web in the European Higher Education Area – an exploration
of visual web indicators", Scientometrics, vol. 74, no. 2, pp 295-308,
2008.
[12] D. A. Batallas and A. A. Yassine, "Information Leaders in Product
development Organizational Networks: Social Network Analysis of the
Design Structure Matrix", IEEE Transactions on Engineering
Management, vol. 53, no. 4, pp. 570-582, 2006.
[13] A. Greve and J. W. Salaff, "The development of corporate social capital
in complex innovation processes", Research in the Sociology of
Organizations, vol. 18, pp. 107-134, 2001.
[14] H. Clark and R. Pradeep, "Visual representation of knowledge networks:
A social network analysis of hospitality research domain", International
Journal of Hospitality Management, vol. 27, pp. 302-312, 2008.
[15] K. Carley, "Knowledge acquisition as a social phenomenon",
Instructional Science, vol. 14, no. 3-4, pp. 381-438, 1986.
[16] C. H. Lin, Social capital: A theory of social structure and action.
Cambridge: Cambridge University Press, 2001.
[17] L. C. Freeman, "Centrality in Social Networks: Conceptual
Clarification", Social Networks, vol. 1, pp 215-239, 1979.
[18] V. Batagelj and A. Mrvar, Pajek – Program for Large Network Analysis,
http://vlado.fmf.uni-lj.si/pub/networks/pajek/
[19] Microsoft Academic Search, http://academic.research.microsoft.com/
115
