An Efficient Modified Common Neighbor Approach for Link Prediction in Social Networks

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661, p- ISSN: 2278-8727Volume 12, Issue 3 (Jul. - Aug. 2013), PP 25-34
www.iosrjournals.org
www.iosrjournals.org 25 | Page
An Efficient Modified Common Neighbor Approach for Link
Prediction in Social Networks
1
Abhisek Chaturvedi, 2
Tapodhir Acharjee
1
Dept of Information Technology, Assam University, Silchar-788011, India
2
Dept of Information Technology, Assam University, Silchar-788011, India
Abstract: Link prediction in social networks aims at estimating the likelihood of the appearance of a new link
between two nodes, based on the existing links and the attributes of the nodes. Many methods for link prediction
problem in social networks have been proposed in literature. We especially analyze the shortcomings of
common neighbor leading method. Accordingly we generate a new modified common neighbor approach for
link prediction in social networks. Our approach efficiently works under the integrated analysis of features
along with topological structure of a social network. As a co-authorship network is a true social network, we
have considered the co-authorship networks for verifying the effectiveness of the existing leading methods as
well as our proposed link prediction method. We have implemented the leading methods as well as our proposed
method on two different data domains of co-authorship networks obtained from author lists of papers at five
sections of Physics e-Print arXiv, www.arXiv.org. In the first data domain, the papers in the periods (1994 –
1996) and (1997 – 1999) are taken as the training set and testing set respectively. Similarly the papers in the
periods (2007 – 2009) and (2010 – 2012) are taken as the training set and testing set for the second data
domain. Experimental results show that all the methods are found to perform much better over the random
predictor. Again we find that our modified common neighbor approach outperforms over all the existing leading
methods considered.
Keywords: co-authorship networks, common neighbor, modified common neighbor approach, link prediction,
random predictor
I. Introduction
Liben-Nowell and Kleinberg [4,5] explain link prediction problem as: “Given a snapshot of a social
network at time t, we seek to accurately predict the edges that will be added to the network during the interval
from time t to a given future time t'.”
In our daily lives we often see many complex networks around us (e.g. Internet, WWW, Transportation
Network, Social Network etc.). The social network may be viewed as a graph where all the nodes or actors are
individuals or organizations and links are contacts (interactions, relationships or collaborations) which show the
proximity or similarity among those individuals or organizations. Natural examples of social networks include
the set of all scientists in a particular discipline, with edges joining pairs who have co-authored papers; the set of
all employees in a large company, with edges joining pairs working on a common project; a collection of
business leaders, with edges joining pairs who have served together on a corporate board of directors.
In past social networks have been studied by many researchers for different purposes. The challenging
problem to deal with these networks has been their highly dynamic nature; they grow and change quickly over
time through the addition of new edges which signifies the appearance of new interactions in the underlying
social structure. Understanding the mechanism by which they evolve is a fundamental question that is still not
well understood by us, and it forms the motivation for our work here on link prediction problem.
Recently the link prediction in complex networks has attracted increasing attention from computer scientists and
physicists. Link prediction aims at estimating the likelihood of the appearance of a new link between two nodes,
based on the existing links and the attributes of the nodes. For example, classical information retrieval can be
viewed as predicting missing links between words and documents, and the process of recommending items to
users can be considered as a link prediction problem in the user-item bipartite networks. Link prediction
problem can be categorized into two categories: one is the prediction of existing yet unknown links (e.g.
criminal/terrorists networks) and the other is the prediction of links that may appear in the future of evolving
networks (e.g., co-authorship networks). Here, we focus on second category of link prediction problem.
Many methods for link prediction problem in social networks have been proposed in literature. Here we
focus on existing leading methods and analyze the shortcomings of some of these methods. Accordingly we
Submitted date 14 June 2013 Accepted Date: 19 June 2013

An Efficient Modified Common Neighbor Approach for Link Prediction in Social Networks
have come out a few new methods for link prediction in social networks. Our methods are based on the
integrated analysis of features along with topological structure of social networks. As a co-authorship network is
a true social network, we have considered the co-authorship networks for verifying the effectiveness of the
existing leading as well as our proposed methods. We have implemented the all the methods on two different
data domains of co-authorship networks obtained from author lists of papers at five sections of Physics e-Print
arXiv, www.arXiv.org. In the first data domain, the papers in the periods (1994 – 1996) and (1997 – 1999) are
taken as the training set and testing set respectively. Similarly the papers in the periods (2007 – 2009) and (2010
– 2012) are taken as the training set and testing set for the second data domain. Here for modeling of all of these
co-authorship networks, we have proposed an efficient algorithm. Experimental results show that all the
methods are found to perform much better over the random predictor on both the data domains. Again we find
that our proposed hybrid approach that is based on the common neighbor-wise combination of our two
proposals (modified Common neighbor and modified Adamic/Adar) gives better results than the existing
leading methods considered on both the data domains.
Rest of paper is organized as follows. In Section 2, we will focus on the literature survey related to link
prediction problem. Section 3 finds a short description of existing leading link prediction methods. Our
proposed link prediction methods are furnished in section 4. Section 5 is concerned with the implementation and
results. The paper is concluded in section 6.
II. Literature Survey
With the best of our knowledge, the first work on link prediction problem in social networks was done
by Liben-Nowell and Kleinberg [4]. They tested the predictive power of some proximity metrics, including
Common neighbours, Adamic/Adar, Katz measure. They further worked on this problem in [5] with more
descriptive analysis of same proximity metrics as in [4]. Here in both the works [4][5], their hypothesis was that
link prediction could be performed from topological analysis alone. Some authors in [6] studied various aspects
like modeling, clustering and ranking in co-authorship networks induced by IEEE, ACM and joint IEEE/ACM
digital library conferences. Some other authors in [7] explored learning based approach where the predictors are
learned using a combination of features and then the performance is evaluated. In 2004, Popescul and Ungar
enhanced link prediction of author/document bipartite networks by using clustering [14]. In [15], Zhou and
Scholkopf approached three related graph problems (classification, ranking and link prediction) in a new way. In
[11], Murata and Moriyasu proposed three weighted similarity indices, as variants of the Common Neighbors,
Adamic-Adar and Preferential Attachment indices, respectively. Some authors in [8] measured the performances
of weighted and unweighted versions of Common neighbor, Adamic-Adar and Resource Allocation on real
social, technological and biological networks, and found that sometimes the weighted indices perform even
worse than unweighted indices. Zhou et al. [9, 10] studied nine well-known local similarity indices on six real
networks extracted from disparate fields, as well as proposed two new local indices. More related is the work on
link prediction using multiple structural attributes and research titles by Pavlov- Ichise and Wohlfarth- Ichise
[12, 13]. They introduced structural features of the graph trying to capture the network structure effectively.
Then, they used supervised learning techniques for link prediction. In [17], Huang, Li and Chen investigated the
use of link prediction to improve collaborative filtering in recommender systems. In [16], link prediction was
used to design a system that recommended new academic links for researchers at a computer science conference
and received feedback through a survey. In [18], Zhu used link prediction to determine what web page a user
was next likely to visit in order to improve the navigation and efficiency of a site. A few other link prediction
papers are summarized by Getoor and Diehl in [19].
III. Existing Leading Link Prediction Methods & Analysis
Many methods have been proposed in literature for link prediction in social networks. Some leading
methods [4][5][11] are : Common neighbor, Jaccard coefficient, Adamic/Adar, Katz Measure, Rooted Page
Rank, Low Rank Approximation, Unseen bigram, Clustering. Here we analyze the shortcomings of methods by
making special intuition for common neighbor method as follows:
 The major problem not only with common neighbor approach but with all of other methods is that they just
focus on topological structure alone to predict the links in social networks. There is no node-similarity based
approach is followed so need to be modified.
 Traditional weighted Common neighbor method [11] which is not considered to be better than unweighted
Common neighbor [4-5], just takes simple addition of all of the weights of common nodes‟ links with nodes X
and Y, no clear logic is shown for this addition so need to be modified.
 Bigram approach [4-5] is not truly defined for link prediction in social networks. The actual bigram approach
refers to find semantic relations between any two nodes X and Y (e.g. authors) based on their features‟ selection.
This limitation for link prediction highly motivates us to work on features for prediction in social networks.

 We can find that till now, no one has provided an unsupervised efficient approach for link prediction in social
networks under integrated analysis of features along with topological structure of social networks.
IV. Proposed Link Prediction Method
After analyzing the shortcomings of common neighbor and bigram existing leading methods in section
3, here we provide a modified common neighbor method which works under integrated analysis of features
along with topological structure of social network. Our approach can be well understood under our suggested
normalized technique that is shown below:
w(X,Y) = 2k/(n1+n2) ……………..(1)
Here n1 and n2 denote total number of features of nodes X and Y respectively; and k denotes the number of
common features between any two connected nodes X and Y.
It generates the weights for each link in present structure which lie between 0 and 1.
4.1 Modified Common Neighbor Method
Common neighbors between nodes X and Y play big roles in prediction of link between them. These
common neighbors can be called „brokers‟ or in concept of social network „structural holes‟ which have the
information about all those nodes which are connected with these nodes („brokers‟). The level of information
that these nodes (brokers) have about other nodes connected with them, depends on the strength of their
relationships. So our proposed method concentrates on the strength of links (relationships) of each common
node („broker‟) with nodes X and Y.
The formula is given as follows:
ScoreMCN(f) (X, Y) = 𝑤1 × 𝑤2
𝑛=𝑁
𝑛=0 …………….……………(2)
Here, N- total number of common neighbors and w1, w2 indicate strength of connections of each respective
common node with X and Y.
If we talk more specifically then our method generates the actual contribution score with respect to strength of
links of each common node with X and Y. Here we can write for each common node: 0<= contribution score
<=1.
We can understand with the help of some examples which highly motivate us. Let us take X and Y any two
unconnected nodes and here Z is a common node between them.
1. If w(X,Z)=1 and w(Z,Y)=0.24 then generated score will be 0.24; Here 1 indicates that both X and Z are fully
similar in behavior so there would be similar likes and dislikes. And now if we focus on the strength of link
between Y and Z then we find, it is not too good that means at a great extent both of them are of different
behavior. In other words, we can say that relationship between Z and Y is not good. Now if Z does not like Y
then X will also not like Y at same level because X and Z are fully similar in nature. So the weight of predictable
link will remain 0.24. In this way, we can say that the generated score is a true score that lies between 0 and 1
for each respective common node between X and Y.
2. If w(X, Z)=1 and w(Z,Y)=1 then generated score will be 1. Obviously, persons in pair (X, Z) and in pair (Z, Y)
are of similar behaviors respectively and know well about one another because of full matching of all features.
So X and Y will also be of similar behavior and in this way, strength of predictable link will be very good (say,
1).
3. If w(X, Z)=0.76 and w(Z,Y)=0.0 then generated score will be 0. Obviously, we can say that Z and Y will not
be connected in present structure and can be unknown to one another at all so it cannot be a good idea to make
prediction between X and Y in this case under integrated analysis of features along with topological structure of
networks.
This proposal holds three key intuitions for taking the summation of each product of weights of each common
node‟s links with nodes X and Y:
 If both the links are strong then predicted link will be strong or almost strong (depending on strength of links
how much they are strong). (e.g. if w1=1 and w2=0.95 then score(X,Y)=0.95).
 If both the links are weak then predicted link will be weak or very weak (depending on the strength of links
how much they are weak). (e.g. if w1=0.3 and w2=0.2 then score(X,Y)=0.06).
 If one link is strong and other is weak then predicted link will also be weak or very weak (depending on the
strength of links how much a strong link is strong and a weak link is weak). (e.g. if w1=1 and w2=0.3 then
score(X,Y)=0.3).

V. Implementation and Results
As a co-authorship network is a true social network and many readers are likely to be interested on
such a network for personal as well as scientific reasons, so we consider co-authorship networks for verifying
the effectiveness of the existing as well as our new method.
5.1 Formulation of Link Prediction Problem
Suppose we have a social network G = < 𝑉, 𝐸 > in which each edge e = < 𝑢, 𝑣 > represents an
interaction between u and v that took place at a particular time t (e). We record multiple interactions between u
and v as parallel edges, with potentially different time-stamps. For two times 𝑡 < 𝑡′
, let G[𝑡, 𝑡′
] denote the
subgraph of G consisting of all edges with a time-stamp between 𝑡 and 𝑡′
. Here, then, is a concrete formulation
of the link prediction problem. We choose four times 𝑡0 < 𝑡0
′
< 𝑡1 < 𝑡1
′
, and give an algorithm access to the
network G[𝑡0, 𝑡0
′
]; it must then output a list of edges, not present in G[𝑡0, 𝑡0
′
], that are predicted to appear in the
network G[𝑡1, 𝑡1
′
]. We refer to [𝑡0, 𝑡0
′
] as the training interval and [𝑡1, 𝑡1
′
] as the test interval.
Of course, social networks grow through the addition of nodes as well as edges, and it is not sensible to seek
predictions for edges whose end points are not present in the training interval. Thus in evaluating link prediction
methods, we will generally use two parameters 𝑘𝑡𝑟𝑎𝑖𝑛𝑖𝑛𝑔 and 𝑘𝑡𝑒𝑠𝑡 , and define the set Core to be all nodes
incident to at least 𝑘𝑡𝑟𝑎𝑖𝑛𝑖𝑛𝑔 edges in G[𝑡0, 𝑡0
′
] and at least 𝑘𝑡𝑒𝑠𝑡 edges in G[𝑡1, 𝑡1
′
]. We will then evaluate how
accurately the new edges between elements of Core can be predicted.
5.2 Preprocessing Work on Datasets
5.2.1 Data Source
Five sections of the physics e-Print arXiv at Cornell University‟s digital library website:
www.arxiv.org. ii) Information about Required Datasets: We work on five co-authorship networks: astro-ph
(astrophysics), cond-mat (condensed matter), gr-qc (general relativity and quantum cosmology), hep-ph (high
energy physics-phenomenology), and hep-th (high energy physics- theory), on two different data domains where
in the first data domain, the papers in the periods (1994 – 1996) and (1997 – 1999) are taken as the training set
and testing set respectively. Similarly the papers in the periods (2007 – 2009) and (2010 – 2012) are taken as the
training set and testing set for the second data domain.
5.2.2 Capturing of Data
We simply access the Cornell University‟s digital library website: www.arxiv.org on a web browser
and we find different sections of the physics e-Print arXiv at this site. Then we find out our required five co-
authorship networks and take the raw data for these networks on both the data domains. This raw data of each
network on both the data domains holds papers‟ information (say, paper_id, title, authors and journal if present).
Finally for both data domains, we place raw data of each network for training and test periods in different text
tiles.
5.2.3 Structuring of Data
As the whole raw data is in unstructured format so our next task is to provide a proper structure of these
raw datasets of each network on both the data domains. We convert the raw data in a proper structural format
(here in each text file having data in structured format, each row holds: paper ids, papers‟ titles, authors and
journal if available otherwise null)) for our whole work on two different data domains.
5.2.4 Core Information
In our experiments on the arXiv, we can identify which authors are active throughout the entire period
on the basis of the number of papers published. Thus here we define the set Core to consist of all authors who
have written at least 𝑘𝑡𝑟𝑎𝑖𝑛𝑖𝑛𝑔 = 3 papers during the training period and at least 𝑘𝑡𝑒𝑠𝑡 = 3 papers during the test
period for each network at first of first dada domain (1994-1999) [Table: 1] and then of second domain (2007-
2012) [Table: 2]. Here we choose „3‟ that can give more possibility to stay the authors in testing period also.

Table 1: Core data of five co-authorship networks at first data domain.
Table 2: Core data of five co-authorship networks at second data domain.
5.3 Implementation
Now we describe our implementation setup more specifically. Let us take any one of five co-authorship
networks in the first data domain (see Table 1), and denote the subgraph G[1994,1996] on the training interval
by Gcollab := <A,Eold> and use Enew to denote the set of edges <u,v> such that <u, v>ϵA(indicating core authors),
and u,v co-author a paper during the test interval but not the training interval - these are the new interactions we
are seeking to predict. In a similar way, we can do for the subgraph G[2007,2009] of any one of these five
networks in second data domain(see Table 2).
Now at this subgraph Gcollab := <A,Eold> for each co-authorship network in both the data domains(see
Table 1 & 2), at first we implement all the leading methods described in section 3 and then for implementing our
proposed link prediction method described in section 4, we focus on titles of papers for getting information
about features of authors/scientists (say, attributes). For this work we use Google stop words vocabulary1
and
Porter stemmer algorithm [24]; and generate the informative keywords (stems) for each author. After finding
these keywords for each author, we implement our proposed method on Gcollab under integrated analysis of
features along with topological structure of social networks.
Further we evaluate each link prediction methods as follows: Each method p that we consider outputs a
ranked list 𝐿 𝑝 of pairs in A× 𝐴 − 𝐸𝑜𝑙𝑑 ; these are predicted new collaborations, in decreasing order of
confidence. For our evaluation, we focus on the set Core (see Table: 1 & 2 in section 5), so we define 𝐸𝑛𝑒𝑤
∗
=
𝐸𝑛𝑒𝑤 ∩ (𝐶𝑜𝑟𝑒 × 𝐶𝑜𝑟𝑒) and n =|𝐸𝑛𝑒𝑤
∗
|. (See Table: 3 & 4) Our performance measure for predictor p is then
determined as follows: from the ranked list𝐿 𝑝 , we take the first n pairs in 𝐶𝑜𝑟𝑒 × 𝐶𝑜𝑟𝑒, and determine the size
of the intersection of this set of pairs with the set 𝐸𝑛𝑒𝑤
∗
.
For whole implementation work, we have used the machine with configuration 48 GB RAM, Intel Xeon
Processor 12 Core 2.4 GHz Speed and our programming platform has been NetBeans IDE 7.1.2 on Windows 7
OS.
1
Online SEO Guide:Google Stop Words Vocabulary: http://www.link-assistant.com/seo-stop-words.html
Networks Training period Core data
authors papers edges authors |𝑬 𝒐𝒍𝒅| |𝑬 𝒏𝒆𝒘|
astro-ph 5308 5806 20913 1564 6162 9086
cond-mat 5783 6628 11687 1315 2186 2362
gr-qc 2139 3286 2900 457 488 504
hep-ph 5433 10257 21694 1771 6142 5564
hep-th 5253 9728 7992 1423 2293 2550
Networks Training period Core data
authors papers edges authors |𝑬 𝒐𝒍𝒅| |𝑬 𝒏𝒆𝒘|
astro-ph 28815 33432 504315 11040 280129 368635
cond-mat 32964 30907 166046 7346 51336 60453
gr-qc 6932 9380 34081 1650 6700 6602
hep-ph 10751 16247 90802 3343 23476 26936
hep-th 8909 15213 20502 2798 7011 6913

Table 3: Finding Enew
∗
for evaluating each link predicting method on each of five co-authorship networks in first
data domain
Table 4: Finding Enew
∗
for evaluating each link predicting method on each of five co-authorship networks in
second data domain
5.4 Implementation Results
Now we show the experimental results of existing and proposed link predicting methods at first on
each of five co-authorship networks in first data domain (See Table:5 and Table:6) and then on each of five co-
authorship networks in second data domain (See Table:7 and Table:8). Here for measuring the overall
performances of existing and proposed methods, we sum the number of correct predictions made by respective
method on each of five networks. For better and meaningful evaluation of all link prediction methods‟ quality,
we use as our baseline a random predictor which simple randomly selects pairs of authors who did not
collaborate in the training interval. Here we also avoid the biasness of random predictor‟s results for each
respective network on both the data domains.
5.4.1 Results on First Data Domain (1994-1999)
Results for each of five co-authorship networks on first data domain are as follows:
5.4.1.1 Experimental Results of Existing LP Methods over Random Predictor
Experimental results of existing link prediction methods over random predictor are given in Table 5.
Here we give abbreviations for some methods which are as follows: RP: Random Predictor, CN: Common
neighbor, AA: Adamic/Adar, JC: Jaccard coefficient, KM(KM_1, KM_2, KM_3): Katz measure with =0.0005,
0.005, 0.005 respectively, RPR: Rooted page rank, LRA: Low rank approximation, KC: Katz clustering, UB:
Unseen bigram), wgtd: weighted, unwgtd: unweighted. Here we place Enew in Table 5 for better visualization of
predictors‟ performances. Methods with bold italic entries give better results than other existing leading methods
considered.
Existing LP Methods astro-ph cond-mat gr-qc hep-ph hep-th Total
𝐸 𝑛𝑒𝑤 9086 2362 504 5564 2550 20066
RP 46 5.6 1 13.4 4.6 70.6
CN(unwgtd) 477 74 11 189 91 842
CN(wgtd) 417 81 8 200 101 807
AA(unwgted) 526 87 16 247 93 969
AA(wgted) 247 89 9 182 96 623
JC 435 68 6 249 81 839
KM_1(wgtd) 382 92 7 176 97 754
KM_2(wgtd) 382 91 7 176 97 753
KM_3(wgtd) 375 89 7 174 95 740
KM_1(unwgtd) 447 77 11 186 99 820
KM_2(unwgtd) 447 77 11 186 99 820
KM_3(unwgtd) 409 74 11 179 99 772
SimRank (γ=0.8) 405 40 11 167 72 695
RPR( α=0.15) 435 70 6 209 86 806
RPR( α=0.50) 410 83 9 196 102 700
LRA(1024)+CN 462 70 12 162 75 841
LRA(512)+CN 399 60 12 151 64 686
LRA(256)+CN 297 54 13 101 62 527
KC(:.001,ρ:.10) 382 92 7 175 95 751
KC(:.001,ρ:.15) 385 92 7 177 96 757
KC(:.001,ρ:.20) 386 92 7 183 97 765
KC(:.001,ρ:.25 379 91 7 184 100 661
UB(δ=4)+CN 414 59 8 214 82 777
UB(δ=5)+CN 437 60 4 215 74 790
𝑬 𝒏𝒆𝒘
∗
= 𝑬 𝒏𝒆𝒘 ∩ 𝑪𝒐𝒓𝒆 ×
𝑪𝒐𝒓𝒆 astro-ph cond-mat gr-qc hep-ph hep-th
𝑪𝒐𝒓𝒆 × 𝑪𝒐𝒓𝒆 59819039 25793119 1119050 5275394 3384699
𝑬 𝒏𝒆𝒘 368635 60453 6602 26936 6913
𝑬 𝒏𝒆𝒘
∗
246524 40192 3395 19426 3911

5.4.1.2 Experimental Results of Proposed LP Methods over Random Predictor
Experimental results of proposed link prediction method over random predictor are given in Table 6.
Here we give abbreviation for Modified Common Neighbor as MCN. Similar to Table 5, here also we place
Enew for better visualization of our proposed link prediction method‟s performance.
Table 6: Performance of proposed link predicting method over random predictor
5.4.2 Results on Second Data Domain (2007-2012)
Results for each of five co-authorship networks on second data domain are as follows:
5.4.2.1 Experimental Results of Existing LP Methods over Random Predictor
Experimental results of existing link prediction methods over random predictor are given in Table 7.
Here we give abbreviations for some methods which are as follows: RP: Random Predictor, CN: Common
neighbor, AA: Adamic/Adar, JC: Jaccard coefficient, KM (KM_1, KM_2, KM_3): Katz measure with
=0.0005, 0.005, 0.005 respectively, RPR: Rooted page rank, LRA: Low rank approximation, KC: Katz
clustering, UB: Unseen bigram), wgtd: weighted, unwgtd: unweighted. Here we place Enew in Table for better
visualization of predictors‟ performances. Methods with bold italic entries give better results than other existing
methods.
Table 7: Performances of existing link predictors over random predictor
𝑬 𝒏𝒆𝒘
∗
= 𝑬 𝒏𝒆𝒘
∩ 𝑪𝒐𝒓𝒆 × 𝑪𝒐𝒓𝒆
astro-ph cond-mat gr-qc hep-ph hep-th
𝑪𝒐𝒓𝒆 × 𝑪𝒐𝒓𝒆 11,45241 704080 69637 1392486 784592
𝑬 𝒏𝒆𝒘 9086 2362 504 5564 2550
𝑬 𝒏𝒆𝒘
∗ 5595 1207 186 3125 1373
Proposed LP
Methods
astro-ph cond-mat gr-qc hep-ph hep-th Total
𝐸 𝑛𝑒𝑤 9086 2362 504 5564 2550 20066
RP 46 5.6 1 13.4 4.6 70.6
MCN 513 100 13 285 111 1022
Existing LP Methods astro-ph cond-mat gr-qc hep-ph hep-th Total
𝐸 𝑛𝑒𝑤 368635 60453 6602 26936 6913 469539
RP 1519 94.2 17.2 94.4 8.4 1733.2
CN(unwgtd) 25752 4114 352 1443 196 31857
CN(wgted) 20093 2876 276 1337 221 24803
AA(unwgtd) 26706 4295 393 1659 227 33280
AA(wgted) 21104 2751 229 1277 220 25581
JC 25207 2252 284 1293 188 29224
KM(wgted) 20113 2890 280 1332 223 24838
KM(wgted) 20101 2888 280 1339 222 24830
KM(wgted) 20085 2881 276 1328 222 24792
KM(unwgted) 24167 3905 330 1205 191 29798
KM(unwgted) 23993 3776 310 1176 190 29355
KM(unwgted) 23908 3442 33 698 152 28233
SimRank γ=0.8 23124 3805 183 1287 192 28591
RPR( α=0.15) 22610 3901 153 1198 151 28013
RPR( α=0.50) 21008 4103 170 1177 143 26601
LRA(1024)+CN 25641 4131 315 1219 130 31436
LRA(512)+CN 24890 4043 271 992 101 30297
LRA(256)+CN 23078 4053 352 899 100 28482
KC(:.001,ρ:.10 20113 2890 283 1332 222 24840
KC :.001,ρ:.15 20317 2932 287 1341 223 25100

5.4.2.2 Experimental Results of Proposed LP Method over Random Predictor
Experimental results of proposed link prediction method over random predictor are given in Table 8.
Here we give abbreviation for Modified Common Neighbor as MCN. Here also we have place Enew for better
visualization of our proposed link prediction method‟s performance.
Table 8: Performance of proposed link predicting method over random predictor
5.4.3 Results based on Execution Time
Here we show the results for execution time taken by our proposed method over existing common
neighbor method. Here these results are shown over gr-qc network in first data domain (1994-1999). In similar
way, we can show for other networks also. Here our intuition in finding the execution time has been to check
whether our method is efficient with respect to execution time over existing CN method or not (See Table 9).
Table 9: Execution time analysis, Here, CN: Common neighbor, MCN: Modified Common Neighbor.
Figure 1: Correct Predictions vs Existing methods and proposed approach on First Data Domain (1994-99).
Here, RP: Random Predictor, CN: Common Neighbor, JC: Jaccard Coefficient, AA: Adamic/Adar, KM: Katz
Measure with =0.0005, RPR: Rooted Page Rank with α=0.15, LRA: Low Rank Approximation with rank-
1024, KC: Katz Clustering with ρ=0.20, UB: Unseen Bigram with δ=4, MCN: Modified Common Neighbor.
KC :.001,ρ:.20 20319 2932 287 1345 226 25109
KC :.001,ρ:.25 20321 2935 284 1349 223 25112
UB(δ=4)+CN 21146 2094 345 1399 189 25173
UB(δ=5)+CN 21713 2368 334 1354 183 25952
Proposed LP
Methods
astro-ph cond-mat gr-qc hep-ph hep-th Total
𝐸 𝑛𝑒𝑤 368635 60453 6602 26936 6913 469539
RP 1519 94.2 17.2 94.4 8.4 1733.2
MCN 26728 4423 362 1802 251 33566
Serial no. Existing Methods Proposed Method Execution Time(sec)
1. CN(unwgted) - 2.746
2. - MCN 10.591
0
5000
10000
15000
20000
25000
30000
35000

Figure 2: Correct Predictions vs Existing methods and proposed approach on Second Data Domain (2007-12).
Here, RP: Random Predictor, CN: Common Neighbor, JC: Jaccard Coefficient, AA: Adamic/Adar, KM: Katz
Measure with =0.0005, RPR: Rooted Page Rank with α=0.15, LRA: Low Rank Approximation with rank-
1024, KC: Katz Clustering with ρ=0.25, UB: Unseen Bigram with δ=5, MCN: Modified Common Neighbor.
VI. Conclusions
Experimental results show that all the existing leading methods considered and our proposed method
are found to perform much better over the random predictor on both the data domains. Again we find that our
approach that works under integrated analysis of features along with topological structure of social networks,
gives better results than the existing leading methods considered on both the data domains.
Our suggestions for further work are as follows:
1. Here in our work, we focus on titles of papers to find the required features of authors. Anyone else can
include other semantic information (e.g. abstracts, journals etc.) and can improve the performance of our
proposed link prediction method.
2. Related to core information on each co-authorship network, we focus only on those authors who write at least
three papers in training and testing periods respectively. We can do our task by taking authors with at least two
papers or greater than three papers on large networks.
3. As our method has been time-consuming than existing leading methods considered (See Table 9 in section 5);
so need to be modified for quick predictions on very large datasets.
4. Here we focus on co-authorship networks as these networks are true social networks. We can make an
intuition for applying our method on other social networks like friendship networks, employees‟ networks in
organizations.
References
[1] Salton G. and McGill M. J., Introduction to Model Information Retrieval (McGraw-Hill, Auckland) 1983.
[2] Zhou T., Ren J., Medo M. and Zhang Y.C., Bipartite network projection and personnel recommendation, Phys.Rev. E, 76 (2007)
046115.
[3] Zhou T., Jiang L.L., Su R.Q. and Zhang Y.C., Effect of initial configuration on network-based recommendation, EPL 81 (2008)
58004.
[4] Liben-Nowell, D., Kleinberg, J.: The link prediction problem for social networks. In: Proceedings of the twelfth international
conference on Information and Knowledge Management, pp. 556-559.ACM Press, New York (2003).
[5] Liben-Nowell D. and Kleinberg J., J. Am. Soc. Inf.Sci. & Technol., 58 (2007) 1019.
[6] Xiaoming Liu, Johan Bollen, Michael Nelson, Herbert Van De Sompel. Co-authorship networks in digital library research
community, Journal of Information Processing and Management, volume 41, issue 6, 2005.
[7] Mohammad Al Hasan, Vineet Chaoji, Saeed Salem, Mohammed Zaki. Link Prediction Using Supervised Learning, In Proc. of
SDM 06 workshop on Link Analysis, Counterterrorism and Security, 2006.
[8] Linyuan Lšu and Tao Zhou, Link prediction in weighted networks: The role of weak ties, EPL, 89 (2010) 18001.
[9] Zhou T., Lšu L. and Zhang Y.-C., Eur. Phys. J. B, 71(2009) 623.
[10] Lšu L., Jin C.-H. and Zhou T., Phys. Rev. E, 80 (2009)046122.
[11] Murata T. and Moriyasu S., Link prediction of social networks based on weighted proximity measures, in Proceedings of the
IEEE/WIC/ACM International Conference on Web Intelligence (ACM Press, New York) 2007.
[12] Pavlov, M., Ichise, R.: Finding experts by link prediction in co-authorship networks. In: Proceedings of the 2nd International
Workshop on Finding Experts on the Web with Semantics (2007).
0
200
400
600
800
1000
1200
RP CN JC AA KM RPRSimRankLRA+CN KC UBMCN

[13] Wohlfarth T., Ichise, R.: Semantic and Event-Based Approach for Link Prediction, in Proceedings of the7th International
Conference on Practical aspects of Knowledge Management 50-61 (2008).
[14] Popescul, A and Ungar, LH 2004, 'Cluster-based Concept Invention for Statistical Relational Learning', Proceedings of Conference
Knowledge Discovery and Data Mining (KDD-2004), 22-25 August 2004, viewed 12 June 2006.
[15] Zhou, D & Scholkopf, B 2004, 'A regularization framework for learning from graph data', Proceedings of Workshop on Statistical
Relational Learning at International Conference on Machine Learning, Banff, viewed 12 June 2006.
[16] Farrell, S, Campbell, C & Myagmar, S 2005, 'Relescope: An Experiment in Accelerating Relationships', paper presented at
Conference on Human Factors in Computing Systems, 2-7 April 2005, Portland.
[17] Huang Z., Li X. and Chen H., Link prediction approach to collaborative filtering, in Proceedings of the 5th ACM/IEEE-CS Joint
Conference on Digital Libraries(ACM Press, New York) 2005.
[18] Zhu, J 2003, 'Mining Web Site Link Structures for Adaptive Web Site Navigation and Search', PhD thesis, University of Ulster,
viewed 12 June 2006.
[19] Getoor and C. P. Diehl, “Link mining: a survey,” SIGKDD Explor. Newsletter, vol. 7, no. 2, pp. 3–12, 2005.
[20] Leo Katz. A new status index derived from sociometric analysis. Psychometrika, 18(1):39-43, March 1953.
[21] Han Hee Song, Tae Won Cho, Vacha Dave, Yin Zhang, Lili Qiu: Scalable proximity estimation and link prediction in online social
networks. Internet Measurement Conference 2009: 322-335.
[22] Sergey Brin and Lawrence Page, The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN
Systems, 30(1-7):107-117, 1998.
[23] Glen Jeh and Jennifer Widom, SimRank: A measure of structural-context similarity, In Proceedings of the ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pages 271-279, July 2002.
[24] M.F. Porter, An algorithm for suffix stripping, Program, 14(3) pp 130−137, 1980.

An Efficient Modified Common Neighbor Approach for Link Prediction in Social Networks

More Related Content

What's hot(18)

Viewers also liked(20)

Similar to An Efficient Modified Common Neighbor Approach for Link Prediction in Social Networks(20)

More from IOSR Journals(20)

Recently uploaded(20)

An Efficient Modified Common Neighbor Approach for Link Prediction in Social Networks