Community Detection in Facebook with Outlier

Recognition
Htwe Nu Win Khin Thidar Lynn
University of Computer Studies, Mandalay Faculty of Information Science
Myanmar University of Computer Studies, Mandalay
htwenuwin99@gmail.com Myanmar
lynnthidar@gmail.com

Abstract² Communities among users play the popular role for
days of Social Network and the presence of groups of nodes that
are high tightly connected with each other than with less links
connected to nodes of different groups. So community detection
algorithms are come to be the key to detect the user who are
interact with each other in social media. However, there are still
challenges in considering of some nodes have no any common
node within the same group as well as some nodes have no any
link to the other node. It can be used similarity measure based on
neighborhood overlapping of nodes to organize communities and
to identify outliers which cannot be grouped into any of the
communities. In this paper, we detect communities and outliers
from Edge Structure with neighborhood overlap by using nodes
similarity. The result implies the best quality with modularity
measurement which leads to more accurate communities as well
as improved their density after removing outliers in the network
structure.
Keywords-Social Network; community; outlier; modularity;
density
This paper explores the use of neighborhood overlapping
by using vertex similarity method for detecting outlier and
significant community. The heart of this approach is to
represent the underlying dataset as an undirected graph, where
a user refers to each node and friendship between two users
represents each edge. Before we measure the similarity among
neighborhood overlap, finding seed node by using the degree
centrality is necessary which is designed to find nodes that are
PRVW ³FHQWUDO´ WR WKH JUDSK :H operate similarity from the
most centrality node and its neighborhood nodes. The values of
zero similarity are then used to identify as outliers. To
illustrate, consider graph of Figure 1, which consists of 23
nodes. Upon applying the detecting outliers method, the
communities which rounded with red circles are group
together, some nodes are saturated with outside them and node
x is saturated lonely. It can be seen clearly the significant
communities, outliers and isolated node in this toy example.

I. INTRODUCTION m Communities
Outlier
Social networks are naturally modeled as graphs, which we
l
sometimes represent it as a social graph. The entities are
nodes, and an edge that connects two nodes if the nodes are e
g h
interacted by the relationship that forms the network. For a
Facebook friendship graph which we will use in our work,
social graphs are undirected and unweighted. The
communities of complex networks are groups of nodes which c
b j
are high tightly connected with nodes of the same group other f i
than with less links connected with nodes of different groups
[22]. Communities may be groups of friendship in social d
networks [1], sets of web pages concerning with the same o
topic and groups of cells with similar functions. While k
Community

identifying the communities in graphs, nodes which cannot Outlier

q
Community

groups with any communities and need not be necessary v u p

group, will be identified as outliers. The early observer of
outlier is Hawkins [4]. He said that ³$Q RXWOLHU LV DQ r s
observation that deviates so much from other observations as w
to arouse suspicion that it was generated by a different
PHFKDQLVP´ Our proposed system is adopted from this Isolated x Outlier t
definition, nodes which have no any friendship or there are no Node

common friends, so much deviate and saturated with

individual lonely is defined outlier. Figure 1. Example of Communities and Outliers

978-1-5090-5504-3/17/$31.00 ©2017 IEEE 155

SNPD 2017, June 26-28, 2017, Kanazawa, Japan
 The rest of paper is organized as follows. We briefly
surveyed related work in section II. In section III, we describe
the background methodology of our work. And then, we
briefly describe about our propose system in section IV. In
section V, we discuss about the experiment and evaluation of
our work. Then, we conclude our work in section VI and talk
about our future idea.
II. RELATED WORK
Many approaches of community and outlier detection
algorithm have been proved over years. Each trend has
efficient and effective in their ways. After the time of Hawkins
[4] definition, many different points of view in outliers were
appeared. Graph based outlier detection were also appeared
with the flow of researchers. Glen Jeh and Jennifer Widom [6]
identified the method of SimRank the proximity measure of
graph that quantified the closeness of two nodes in the graph,
which based on their relationships with other objects and
computed the similarity of the structural context in which the
graph objects occur. Several variants of SimRank were also
proposed by [9], [8] and [21]. On the other hand, Keith
Henderson et al., [10] proposed the one could use to compute
various measures associated with the nodes in the given graph
structure, dyads, triads, egonets, communities, as well as the
global graph structure. Leman Akoglu et al., [13] proposed
OddBall which is an algorithm to detect anomalous nodes in
weighted graphs. Detecting anomalous sub-graphs using
variants of the Minimum Description Length (MDL) principle
was proposed by Caleb C. Noble and Diane J. Cook, [2]. H. D.
K. Moonesinghe and Pang-Ning Tan [7] proposed Outrank
method which used the MDL principle as well as other
probabilistic measures to detect several types of anomalies
(e.g. unexpected/missing nodes/edges). OutRank and
LOADED [3, 17] used similarity graphs of objects to detect
outliers. J.Y. Chen et al., [10] used MDL to spot anomalous
edges. Jimeng Sun et al., [11] used proximity and random
walks, to assess the normality of nodes in bipartite graphs.
Meng Wang et al., proposed the method of detecting outliers
with community by using minimum valid size (mvs). If the
minimum size is 2, it will be chosen the single node and
marked it as outlier [18]. In paper [5] and [24], the method of
community detection based on link and relationship are
determined. In contrast to the above, this paper is used
undirected and unweighted graph data based on overlapping
neighborhood. It is explicitly focused on edge structure to
detect outliers and significant communities with nodes
similarity.
III. BACKGROUND METHODOLOGY
A. Graph
We consider basic definitions of a given network
represented as a directed graph or undirected graph. Facebook
friendship network which we will use in our work here, a
graph may be undirected, meaning that there is no distinction
between the two vertices associated with each edge.
The notation of a graph G = (V,E) consists of two sets V
and E. The elements of V = {v1, v2, . . . , vN} are the nodes or
156
vertices of the graph G where each vertex vi is associated with
the instance xi from the input data X and the cardinality of |V |
is N. The elements of E = {e1, e2, . . . , eM} are links or edges
between nodes and the cardinality of |E| is M. An edge
connecting the vertices vi and vj is denoted by eij [19].
B. Node Degree and Its Neighborhood
In network G, the degree of any node ݅ is the number of
nodes adjacent to ݅. Generally, the more degree that the node
has, the more important it will be. Two vertices ‫ ݒ‬and ‫ ݑ‬are
called neighbors, if they are connected by an edge. Let ʒ௜ be
the neighborhood of vertex ݅ in a graph, i.e., the set of vertices
that are directly connected to ݅ via an edge [14].
C. Communities
The communities of social network are groups of nodes,
with more links connecting nodes of the same group and
comparatively less links connecting nodes of different groups.
Communities may be groups of related individuals in social
networks. Identifying communities in a network can provide
valuable information about the structural properties of the
network, the interactions among nodes in the communities,
and the role of the nodes in each community [22].
In an undirected graph ‫ܩ‬ሺܸǡ ‫ܧ‬ሻǡ where the total
number of node, ȁܸȁ ൌ ݊ and total number of edges,ȁ‫ܧ‬ȁ ൌ ݉
are defined. We can identify set of communities such that
‫ ݏ݉݋ܥ‬ൌ ሼܸͳƍ ǡ ܸʹƍǡ ǥ ǥ Ǥ Ǥ ǡ ܸܿ݊ƍሽ where ‫ڂ‬௖௡
௜ୀଵ ܸ௜ ƍ ‫ܸ ك‬௜  and ܿ݊ is
the total number of communities Coms should satisfy,
ܸ݅ ƍ ‫ ݆ܸځ‬ƍ ൌ ߶ [18].
D. Vertex Centrality
The importance of a node is determined by the
number of nodes adjacent to it. The larger the degree of node,
the more important the node is. Those high-degree nodes
naturally have more impact are considered to be more
important. The degree centrality is defined as
‫ܥ‬஽ ሺ‫ݒ‬௜ ሻ ൌ ݀௜ ൌ σࣼ ‫ܣ‬௜௝ (1)
When one needs to compare two nodes in different
networks, a normalized degree centrality should be used,
‫ܥ‬஽ Ԣሺ‫ݒ‬௜ ሻ ൌ ݀௜ Ȁሺ݊ െ ͳሻ. (2)
Here, n is the number of nodes in a network. It is the
proportion of nodes that are adjacent to node ‫ݒ‬௜ [14].
E. Outliers
According to Hawkins [4], outliers can be defined as
follows: ³$Q RXWOLHU LV DQ REVHUYDWLRQ WKDW GHYLDWHV VR PXFK
from other observations as to arouse suspicion that it was
JHQHUDWHG E\ D GLIIHUHQW PHFKDQLVP´ 0RVW RXWOLHU GHWHFWLRQ
VFKHPHV DGRSW +DZNLQ¶V GHILQLWLRQ of outliers and thus
assume in our system is that outliers are nodes which have
zero values of similarity measure in graph. Each node in a
graph cannot be grouped into any of the communities is called
outlier. As such, these outliers can be easily detected by
existing distance or density based algorithms. However, in this
paper we focus on outliers that might be concentrated in
certain regions. We take edge structure based approach in
graph to solve this problem. Outliers and communities can be
defined by:
 ܱ‫ ݏݐݑ‬ൌ ሼ‫ݒ‬ȁ‫ܸ א ݒ‬ǡ ൓‫ ܸ݅׌‬ᇱ ‫ ܸ݅ א ݒ ר ݏ݉݋ܥ א‬ᇱ ሽ ൌ ܸ െ ‫ڂ‬௖௡
௜ୀଵ ܸ݅Ԣ
(3)
Outliers directly identified by getting nodes from small
communities [18]:
‫ ݏ݉݋ܥ‬ራ ܱ‫ ݏݐݑ‬ൌ ‫ݒ‬
‫ ݏ݉݋ܥ‬ሩ ܱ‫ ݏݐݑ‬ൌ Ͳ
Where, ‫ ݏ݉݋ܥ‬is the community and ܱ‫ ݏݐݑ‬is indicated as
outlier.
F. Isolated Node
Some users who used social network, any friendship are
not belonging to them can be included in social network. Such
loneliness users are represented as isolated nodes in this paper.
Since, there are no neighbor nodes around them as well as
their edge structure could not be considered.
Before calculating the seed node, we find them and define
as outlier firstly. As shown in Fig1, node x is isolated and it can
be made as outlier.
G. Vertex Similarity
It can be assumed that communities are groups of vertices
similar to each other. We can compute the similarity between
each pair of vertices after searching seed nodes. Most existing
similarity method are based on the measurement of distance
called Euclidean, Manhattan and etc., Although, to considered
the similarity between selected node and is neighborhood,
Jaccard Similarity is more convenient in this work which we
will measure the similarity based on the neighborhood overlap
of seed nodes. Let ܰ௜  denote the neighbors of node ‫ݒ‬௜ . Given
a link e(‫ݒ‬௜ ǡ ‫ݒ‬௝ ), the neighborhood overlap is defined as
௡௨௠௕௘௥௢௙௦௛௔௥௘ௗ௙௥௜௘௡ௗ௦௢௙௕௢௧௛௩೔ ௔௡ௗ௩ೕ
ܱ‫݌݈ܽݎ݁ݒ‬ሺ‫݅ݒ‬ǡ ‫݆ݒ‬ሻ ൌ
௡௨௠௕௘௥௢௙௙௥௜௘௡ௗ௦௪௛௢௔௥௘௔ௗ௝௔௖௘௡௧௧௢௔௧௟௘௔௦௧௩೔ ௔௡ௗ௩ೕ
ȁே೔ ‫ ځ‬ேೕ ȁ
ൌ หே
೔ ‫׫‬ேೕ หିଶ
We have െʹin the denominator just to exclude ‫ݒ‬௜ ܽ݊݀‫ݒ‬௝
from the set ܰ௜ ‫ܰ ׫‬௝ . If there are no overlap vertices in any
two ܰ௜ ܽ݊݀ܰ௝ means หܰ௜ ‫ܰ ځ‬௝ ห ൌ ‫ ׎‬, we can identify ܰ௝
disconnect withܰ௜ . The neighborhood overlap is positively
correlated with the total number of minute spent by two
persons in a telecommunication network [14]. For instance, in
Figure 2, two terminal nodes, d and e have only one edge
between them. They have no any other neighborhood overlap.
Therefore, their overlap value is 0 as shown in (5). Again, the
neighborhood overlap of other two terminal nodes, e and f is
node h. We can see their overlap value in (6).
e f
a
b d
g h
c this work. There are 347 nodes, represent as users and 2519
Figure 2. Example of Neighborhood Overlap
ȁሼ ሽȁ ଴
ܱ‫݌݈ܽݎ݁ݒ‬ሺ݀ǡ ݁ሻ ൌ ȁሼ௔ǡ௕ǡ௖ǡௗǡ௘ǡ௙ǡ௚ǡ௛ሽȁିଶ ൌ ൌͲ
଼ିଶ
ȁሼ௛ሽȁ ଵ ଵ
ܱ‫݌݈ܽݎ݁ݒ‬ሺ݁ǡ ݂ሻ ൌ ȁሼ௘ǡ௙ǡ௚ǡ௛ሽȁିଶ ൌ ൌ (6)
ସିଶ ଶ
Assuming like that, our work identify outliers are appeared
with among separated communities.
IV. PROPOSED APPROACH
Network can be represented as graph for detecting outliers
and significant communities. In this paper, we use concept of
undirected graph in which two vertices have common node, if
they share one or more of the same vertices. Let Ƚ௜ be the
neighborhood of vertex ݅ in a network, i.e, the set of vertices
that are directly connected to ‫ ݅ݒ‬via an edge. Then the number
of common friends of ‫ ݅ݒ‬and ‫ ݆ݒ‬is ȁȽ௜ ‫ ת‬Ƚ௜ ȁ . Contrary, nodes
which have no any other node in common or sometime they
can be isolated, we can identify them as outliers, can be given
as ȁȽ௜ ‫ ת‬Ƚ௜ ȁ ൌ Ͳ . In our work, we propose the system to
identify significant community with detecting outlier by using
node similarity depend on this concept.
In this work we use two main methods, firstly the vertex
centrality which is used to determine seed nodes with the
highest degree or the largest number of neighborhood and then
use the method of neighborhood overlap based on vertex
Similarity. In this work, we used the following properties:
x Node which has the largest degree is determined as
seed node.
x Determine seed node by using Vertex Centrality
Method.
x Calculate Vertex Similarity among nodes that are
linked to seed node.
x Nodes that belong to the same communities are likely
to be more links connected to each other.
x Each community is likely to be fewer links outside
the rest of the graph.
(4) x The isolated individual nodes are stand as outliers.
x Nodes in the same community are likely to share
common neighborhood node.
x If nodes do not have common neighborhood, they are
acted as outliers.
V. EXPERIMENTS
A. Description of Datasets
In this paper, real undirected networks, and the fraction of
SNAP Social circles: Facebook1 Dataset is used. This dataset
FRQVLVWVRILQIRUPDWLRQRIIULHQGV¶FLUFOHVIURP)DFHERRNIRU
instance, the profiles for each user (features), circles and
network. ,Q WKLV 'DWDVHW VWDWLVWLFV ³QRGHV´ UHSUHVHQWV WKH
QXPEHU RI IULHQGV ³(GJHV´ UHSUHVHQWV WKH QXPEHU RI
friendship in the network. Thus, 4,039 nodes and 88,234 edges
are totally included in this dataset. Since this Facebook graph
separated 10 files which have same features are cluster
together in each. :H XVH HGJH VWUXFWXUH RI ³(GJHV´ ILOH LQ
edges for their friendship information as shown in Table I.
(5)
1
https://snap.stanford.edu/data/egonets-Facebook.html
157
TABLE I. NUMBER OF NODES AND EDGES IN REAL FACEBOOK SOCIAL ϭ͘Ϯ
NETWORK DATASET

Network Node Edge ϭ

SNAP 347 2519 Ϭ͘ϴ

'DWDVHW³(GJH´)LOH
Ϭ͘ϲ
B. Results and Evaluation
The communities and outliers result of our evaluation are Ϭ͘ϰ
shown in Table II, Table III and Table IV. Firstly, they are
removed the isolated node which are included in this network Ϭ͘Ϯ
(represented as graph dataset) since it is based on edge
structure to calculate the outlier. In this paper, there are such Ϭ
14 isolated nodes that they have no any friendship with other Ϭ ϭ Ϯ ϯ ϰ ϱ ϲ ϳ ϴ ϵ ϭϬ ϭϭ ϭϮ ϭϯ ϭϰ ϭϱ ϭϲ ϭϳ ϭϴ ϭϵ ϮϬ Ϯϭ
node. Therefore, they are calculated the total retrieve outliers
/ŶƚƌĂͺůƵƐƚĞƌĞŶƐŝƚLJ /ŶƚĞƌͺůƵƐƚĞƌĞŶƐŝƚLJ
without these isolated nodes. Therefore, 333 nodes are based
to make the similarity among the neighborhood of dataset. Figure 3. Density Measurement of Communities
There are 22 communities as output with removing outliers
and isolated nodes. In Table II we can see the 22 communities, We got the highest density 1 into community 13, 17, 20
84 outliers and 14 isolated nodes. and 21, so we can see them as clique sub_graph in which
every node have edge incident to every other nodes within the
TABLE II. NUMBER OF COMMUNITIES, OUTLIER AND ISOLATED NODE same community. On the other hand, the density of
Network Community Outlier Isolated Node community 1 is lower than the other communities. However
its density measurement value is satisfied with the rule of
SNAP 22 84 14
density for community.
'DWDVHW³(GJH´
File
The value of modularity for communities without
removing outliers is 0.26 by analyzing in Gephi2. Contrary,
with removing outliers, the modularity value of our
communities is o.33. Since outliers are removed the
modularity value of community structure are increased as
shown in Table III. Here, let us talk about the brief definition
of modularity. In [20],[15],[16],[20], modularity Q is a
parameter which is used to define the quality measure of the
community. The Q value is between 0 and 1 and the real
network modularity function value is generally between 0.3
and 0.7. Community structure is more obvious with the greater
value of modularity. By using our removing outlier method,
the modularity value is increased obviously.
TABLE III. NUMBER OF COMMUNITIES AND MODULARITY VALUE
BEFORE AND AFTER REMOVE OUTLIER
Status Q(Modularity) for Community For SNAP
If the fraction of edges linking nodes with different
attribute values is significantly less than the expected
probability, then there is evidence of correlation [14].
Therefore, our approach offered the significant result of
community as shown in Table IV.
Therefore, it can be decided that considering detecting
outlier is effective to get the significant community in social
network.
VI. CONCLUSION
This paper proposed the method for detecting outlier with
grouping the community of Facebook social network by using
neighborhood overlap with vertex similarity method. We
described the detail computation of the proposed system which
was defined outliers with have no common node and was
grouped community with common node. Previously, many
methods to detect community, they only considered the
structure of community without outlier. However, in time of
social network different type of users can be played colorful.
So, thinking of outliers becomes the essential part in method
SNAP 'DWDVHW³(GJH´)LOH of community structure. Therefore, we proposed outlier
'DWDVHW³(GJH´File
Before Remove 0.26 63 detecting method by using edge structure to get the effective
Outlier community. In this paper, we based on edge structure of
After Remove Outlier 0.33 22 unweighted and undirected graph to detect the outlier and
significant communities without thinking any information of
the structure.
As shown in Figure 3, the values of density are satisfied
We used real network dataset from SNAP with the
with the rule of both measurements as well as each
modularity quality function to show the significant result of
intral_cluster density is larger than its corresponding
our method. After that, we used the measurement of
inter_cluster density.
community density to define how much our community is
good and significant. By using removing outlier method with
detecting community the density of our method¶V UHVXOW LV
2
https://gephi.org

158
good in measurement. The drawback of our method is only Conference on Knowledge Discovery and Data Mining (SIGKDD),
Edmonton, Alberta, 2002, pages 538±543.
based on edge structure without considered any information
[7] H. D. K. Moonesinghe and Pang-Ning Tan. ³Outrank: a graph-based
such as profile of user. So, our future will be used feature of outlier detection framework using random walk´ International Journal
users to detect effective community. on Artificial Intelligence Tools, 17(1), 2008.
[8] Hung-Hsuan Chen and C. Lee Giles, ³ASCOS: an asymmetric network
TABLE IV. THE EVIDENCE OF AVERAGE CORRELATION AND NON- structure context similarity measure´ In IEEE/ACM International
CORRELATION AMONG COMMUNITIES AFTER REMOVEING OUTLIER Conference on Advances in Social Networks Analysis and Mining
(ASONAM), Niagara Falls, Canada, 2013.
Community Non-Correlation Correlation
[9] Ioannis Antonellis, Hector Garcia-Molina, and Chi-Chao Chang,
³Simrank++: Query rewriting through link analysis of the click graph´
0 15.73% 1.81%
In Proceedings of the 34nd International Conference on Very Large Data
1 39.64% 2.33% Bases (VLDB), Auckland, New Zealand, 2008, pages 408±421.
[10] J.Y. Chen, O. R. Zaiane, R. Goebel, ³Detecting Communities in Large
2 29.76% 1.36% Networks by Iterative Local Expansion´ International Conference on
3 42.77% 1.06% Computational Aspects of Social Networks, 2009.
[11] Jimeng Sun, Huiming Qu, Deepayan Chakrabarti, and Christos
4 39.76% 0.63% Faloutsos, ³Neighborhood formation and anomaly detection in bipartite
5 29.76% 0.88% graph´ICDM, November 27-30 2005.
[12] Keith Henderson, Tina Eliassi-Rad, Christos Faloutsos, Leman Akoglu,
6 42.77% 3.36% Lei Li, Koji Maruhashi, B. Aditya Prakash, and Hanghang Tong,
7 39.64% 0.71% ³Metricforensics: A multi-level approach for mining volatile graphs´ In
Proceedings of the 16th ACM International Conference on Knowledge
8 43.07% 5.46% Discovery and Data Mining (SIGKDD), Washington, DC, 2010, pages
163±172.
9 38.75% 0.19%
[13] Leman Akoglu, Mary McGlohon and Christos Faloutsos , ³Anomaly
10 41.51% 0.07% Detection in Large Graphs´, November 2009, CMU-CS-09-173.
[14] /HL 7DQJ DQG +XDQ /LX ³&RPPXQLW\ 'HWHFWLRQ DQG 0LQLQJ LQ 6RFLDO
11 41.51% 0.07% 0HGLD´ ,6%1 $ 3XEOLFDWLRQ LQ WKH 0RUJDQ 
12 42.77% 1.59% Claypool Publishers series, 2010.
[15] 0 ( - 1HZPDQ DQG 0 *LUYDQ ³)LQGLQJ DQG HYDOXDWLQJ FRPPXQLW\
13 42.64% 1.65% structure in networks´Phys. Rev. E, vol. 69, p. 026113, Feb 2004.
14 43.07% 0.38% [16] 0( - 1HZPDQ ³0RGXODULW\ DQG community structure in networks´,
Proceedings of the National Academy of Sciences, vol. 103, no. 23,
15 41.28% 0.04% pp.8577±8582, 2006.
16 38.75% 0.29% [17] 0*LUYDQ 0(-1HZPDQ ³Community structure in social and
biological networks´, Proc.Natl. Acad. Sci. USA 99, 7821-7826, 2002.
17 42.64% 0.07% [18] Meng Wang, Chaokun Wang, Jeffrey Xu Yu and Jun Zhang,
18 43.07% 0.16% ´Community Detection in Social Networks:An Indepth Benchmarking
Study with a ProcedureOriented Framework´ 41st International
19 38.75% 1.12% Conference on Very Large Data Bases, August 31 st September 4th 2015.
20 38.75% 0.33% [19] 0LFKHO 3ODQWLH¶ DQG 0LFKHO &UDPSHV ³Survey on Social Community
Detection´, Social Media Retrieval, Springer Publishers, Computer
21 38.75% 0.31% Communications and Networks, 978-1-4471-4554-7, pp.65-85, 2013.
[20] 0LQJPLQJ &KHQ 7RPP\ 1JX\HQ DQG %ROHVODZ . 6]\PDQVNL  ³2Q
0HDVXULQJ WKH 4XDOLW\ RI D 1HWZRUN &RPPXQLW\ 6WUXFWXUH´
Proceedings of the IEEE Social Computing Conference, Washington
REFERENCES DC, September 8-14, pp. 122-127,2013.
[1] Amol Ghoting, Matthew Eric Otey, and Srinivasan Parthasarathy, [21] Peixiang Zhao, Jiawei Han, and Yizhou Sun. ³P-rank: a comprehensive
³/RDGHGLink-based outlier and anomaly detection in evolving data structural similarity measure over information networks´ In Proceedings
sets´ In ICDM, 2004. of the 18th ACM Conference on Information and Knowledge
[2] CDOHE & 1REOH DQG 'LDQH - &RRN ³Graph-based anomaly detection´ Management (CIKM), Hong Kong, China ACM, 2009, pages 553±562.
In KDD, 2003, pages 631±636. [22] S. Fortunato, ³&RPPXQLW\ 'HWHFFWLRQ LQ *UDSK´ Phys. Rep. 486, 75
[3] Deepayan Chakrabarti, ³$XWRSDUWParameter-free graph partitioning and (2010).
outlier detection´ In PKDD, 2004, pages112±124. [23] William Eberle and Lawrence B. Holder. ³Discovering structural
[4] D. Hawkins, ³Identification of outliers´ Chapman and Hall, 1980. anomalies in graph-based data´ In ICDM Workshops, 2007, pages 393±
398.
[5] Fei Teng, Rongjie Dai, Hongjie Wang and Xiaoliang Fan, ³$FFHOHUDWLQJ
/LQN &RPPXQLW\ 'HWHFWLRQ LQ 6RFLDO 1HWZRUNV´ 2015 International [24] Xiang Li, Yong Tan, Ziyang Zhang, and Qiuli Tong ³Community
Conference on Cloud Computing and Big Data (CCBD) (2015), Detection in Large Social Networks Based on Relationship Density´
Shanghai, China, Nov. 4, 2015 to Nov. 6, 2015, ISBN: 978-1-4673- 2016 IEEE 40th Annual Computer Software and Applications
8350-9, pages 119-126. Conference (COMPSAC) (2016), Atlanta, GA, USA, June 10, 2016 to
June 14, 2016, ISSN: 0730-3157, ISBN: 978-1-4673-8846-7, page 524-
[6] Glen Jeh and Jennifer Widom. ³SimRank: a measure of structural- 533.
context similarity´ In Proceedings of the 8th ACM International

159