ComSim : A bipartite community detection algorithm

using cycle and node’s similarity

Raphael Tackx, Fabien Tarissan, Jean-Loup Guillaume

To cite this version:

Raphael Tackx, Fabien Tarissan, Jean-Loup Guillaume. ComSim : A bipartite community detection
algorithm using cycle and node’s similarity. Complex Networks 2017, Nov 2017, Lyon, France. pp.278-
289, �10.1007/978-3-319-72150-7_23�. �hal-01657093�

HAL Id: hal-01657093

https://hal.science/hal-01657093
Submitted on 6 Dec 2017

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
C OM S IM: A bipartite community detection
algorithm using cycle and node’s similarity

Raphael Tackx, Fabien Tarissan, and Jean-Loup Guillaume

Abstract This study proposes C OM S IM, a new algorithm to detect communities in

bipartite networks. This approach generates a partition of > nodes by relying on
similarity between the nodes in terms of links towards ⊥ nodes. In order to show
the relevance of this approach, we implemented and tested the algorithm on 2 small
datasets equipped with a ground-truth partition of the nodes. It turns out that, com-
pared to 3 baseline algorithms used in the context of bipartite graph, C OM S IM pro-
poses the best communities. In addition, we tested the algorithm on a large scale
network. Results show that C OM S IM has good performances, close in time to Lou-
vain. Besides, a qualitative investigation of the communities detected by C OM S IM
reveals that it proposes more balanced communities.

Key words: Community detection; bipartite graph; social network

1 Introduction

Many complex networks lend themselves to the use of graphs for analyzing and
modelling their structure. Usually, vertices of the graph stand for the nodes of the
network and the edges between vertices stand for (possible) interactions between
nodes of the network. This approach has proven to be useful to identify non triv-
ial properties of the structure of networks in very different contexts, ranging from
computer science (the Internet, peer-to-peer networks, the web), to biology (protein-

Raphael Tackx
Sorbonne Universités, CNRS, LIP6, UMR 7606, e-mail: raphael.tackx@lip6.fr
Fabien Tarissan
Universités Paris-Saclay, CNRS, ISP, cole Normale Suprieure de Paris-Saclay e-mail:
fabien.tarissan@ens-paris-saclay.fr
Jean-Loup Guillaume
University of La Rochelle, L3I e-mail: jean-loup.guillaume@univ-lr.fr

1
2 Raphael Tackx, Fabien Tarissan, and Jean-Loup Guillaume

protein interaction networks, gene regulation networks), social science (friendship

networks, collaboration networks), linguistics, economy, etc. [1, 2, 3, 4, 5, 6, 7].
This abstraction into graphs allows in return to study formally different aspects
of its structure. In this context, one question that has driven a lot of attention in the
past decade is the identification of communities, that is sets of nodes that consti-
tute cohesive groups inside the networks. Although no formal definition has led to a
consensus in the scientific community, one usually assumes that members of a com-
munity should be more connected to each other than with the rest of the network.
To identify such communities, one can rely on human expertise but, in the context
of large-scale networks, the question of identifying automatically such communities
has led to the proposition of several community detection algorithms [8].
It is striking to notice that most algorithms have been designed for graphs con-
taining only one set of nodes. Although useful, such a simple representation is not
particularly close to the real structure of most of real networks. If one considers for
instance an actor network that links actors performing in the same movies [1, 9] or
co-authoring network that links authors publishing together [9, 3], one would rather
relate actors to the movies they performed in and authors to their papers. This obser-
vation led the community to use bipartite graphs instead, i.e. graphs in which nodes
can be divided into two disjoint sets, > (e.g. movies) and ⊥ (e.g. actors), such that
every link connects a node in > to one in ⊥.
In that regard, only few community detection methods have been proposed to
take into account this inherent bipartite complexity of real networks [10, 11, 12].
The usual approach consists instead in projecting first the bipartite structure over one
set of nodes and then applying standard community detection techniques. Although
interesting, it has been shown that this approach suffers from limitations [13].
Our contribution in this paper is to propose a new community detection algorithm
dedicated to bipartite networks, namely C OM S IM (Section 2). This algorithms relies
on a measure of similarity between nodes exploiting the bipartite ties. Then the
algorithm looks for cycles of connections maximizing the similarity between the
nodes, thus defining the core of the communities.
In order to validate our approach, we rely on real dataset and compare the com-
munities generated by our algorithm to baseline methods (Section 3). Results show
that on dataset equipped with ground-truth communities, the communities inferred
by C OM S IM are the closest to the real ones. We also show that C OM S IM obtains
good results when applied on large-scale networks as it produces communities that
are more homogeneous than the other approaches tested in this study.

2 New community detection algorithm: C OM S IM

In this section, we formally presents our detection algorithm devoted to bipartite

graphs. We first recall the necessary definitions (Section 2.1) before describing
C OM S IM algorithm (Section 2.2) and presenting baseline algorithms to which we
compare our approach (Section 2.3).
C OM S IM: A bipartite community detection algorithm using cycle and node’s similarity 3

2.1 Notations

A bipartite graph is defined by a triple B = (>, ⊥, EB ) (see Figure 1 for instance)

where > is the set of top nodes (e.g movies), ⊥ the set of bottom nodes (e.g. actors),
and Eb ⊆ > × ⊥ the set of links between > and ⊥ (that relates for instance the actors
to the movies they perform in).

Fig. 1 Example of a bipartite graph B = (>, ⊥, EB ).

In addition, we define N> (v) = {x ∈ ⊥|(v, x) ∈ Eb } as the set of neighbors of a

node v ∈ >1 and N>2 (v) = N⊥ (N> (v)) as the set of neighbors at distance 2 from v,
that is the set of > nodes that share a ⊥ node with v. Then we denote by d> (v) =
|N> (v)| the degree of a node v ∈ >, d> (v), and d> 2 (v) = |N 2 (v)| its number of
>
neighbors at distance 2.
Compared to unipartite graphs, nodes in a bipartite graph are separated in two
disjoint sets, and the links are always between a node in one set and a node in the
other set. But it is natural to also investigate how nodes from the same set are in
relation. This approach is usually captured by the notion of projection of a bipartite
graph over one of its two sets.

Fig. 2 Example of the weighted >-projection of B using common neighbors as similarity function.

For instance, if one is interested in the >-projection, one can study how > nodes
connect according to their similarity measured by their links towards common ⊥
nodes. Formally, such a similarity is captured by a similarity function θ . This allows

1 We use a similar definition of N⊥ (v) for v ∈ ⊥.

4 Raphael Tackx, Fabien Tarissan, and Jean-Loup Guillaume

to formally define the weighted projected graph G> = (>, θ ) where θ : > × > 7→
R+ . This graph thus indicates the strength of the relations between > nodes. The
>-projection of the bipartite graph in Figure 1 will therefore result in the graph
depicted Figure 2.
Note that in the rest of the paper, we will use the standard common neighbor
function θ (x, y) = |N> (x) ∩ N> (y)|. But this approach easily extends to other simi-
larity functions such as jaccard index [14], resource allocation [15] or adamic-adar
coefficient [16]2 .

2.2 C OM S IM algorithm

Given a bipartite graph B = (>, ⊥, EB ) and a similarity function θ such as common

neighbors, C OM S IM generate a partition of > in two steps. First, it identifies the
core communities, that are groups of > nodes highly similar according to function θ .
As one can see in Algorithm 1 which details this first step, the algorithm gen-
erates a chain of nodes by following out-going links that have the highest weight
according to θ . When this chain reaches a node already considered, it means that
a cycle has been detected in the chain. This cycle then forms the core of a future
community.
On the toy example of Figure 1, it would result in detecting that nodes A and B
form the core of a community, as well as E and F. This is in accordance to Figure 2
which shows that A and B, as well as E and F have the highest weighted links. The
other nodes (C and D) are left in the remaining set K.
The second phase of the algorithm then tries to position the remaining nodes of
K in the existing communities by maximizing the similarity between these nodes
and all the nodes of the core communities.
As described in Algorithm 2, the second step considers all remaining nodes that
are not part of the partition after the first step. For each node x, it identifies the
communities that have at least one link with x. The algorithm then chooses the
community that maximizes the sum of similarities between x and all the nodes of
the community.
On the toy example of Figure 1, and independently of the order in which nodes
C and D are considered during step 2, it would result in affecting node C to the
community A − B (the sum of similarities is 3) and node D to community E − F (the
sum of similarities is 4).
It is worth noticing that because several links can have a similar weight, the
two steps might face several equal options. In that case, the algorithm selects one
option uniformly at random among all possible ones. For this reason, the algorithm
is undeterministic and several runs might end up with different partitions.

2 Depending on the similarity function used, the projection might result in a directed weighted
graph if θ is not symmetric.
C OM S IM: A bipartite community detection algorithm using cycle and node’s similarity 5

Algorithm 1: C OM S IM- first step

Data: a bipartite graph B = (>, ⊥, EB ), a similarity function θ
Result: return a partition P of > nodes and a set K of remaining nodes (for the second step)
P := 0/ // the partition set
T := > // the set of nodes to be considered
x := rand and remove(T ) // random node
V := 0/ // set of nodes currently considered
K := 0/ // set of remaining nodes
while T 6= 0/ do
/* finds a neighbor y ∈ N>2 (x) of x maximizing θ (x, y) */
y := argmaxy∈N 2 (x) θ (x, y)
>
if y ∈ V then
C := cycle(V, y, x) // extract the detected cycle from y to x in V
P.add(C)
K := K ∪ (V −C) // stores nodes not in the cycle C
V := 0/
x := rand and remove(T )
else
if y ∈ T then
V := V ∪ {y}
x := y
T := T − {y}
else
/* y is already part of an element of P, visited nodes are stored */
K := K ∪V
V := 0/
x := rand and remove(T )

return P and K

Algorithm 2: C OM S IM- second step

Data: a bipartite graph B = (>, ⊥, EB ); a partition P; a set K of remaining nodes (from first
step), a similarity function θ .
Result: return a partition P0 of > nodes and unsatisfied nodes R
R := 0/ // Remaining nodes
P0 := P
foreach x ∈ K do
Px := com neigh(x, P) // Find all neighbor communities of x
if Px = 0/ then
R := R ∪ {x}
else
C := argmaxCx ∈Px ∑y∈Cx θ (x, y)
Add x into the partition C of P0

return P0 and R

2.3 Standard approaches

In order to evaluate the relevance of C OM S IM, we will compare the detected com-
munities with the ones of the three baseline detection algorithms described below.
6 Raphael Tackx, Fabien Tarissan, and Jean-Loup Guillaume

Louvain: Louvain algorithm [17] is a greedy algorithm that optimizes a qual-

ity function in order to extract communities from large unipartite networks. It is
commonly used with modularity [18] which measures the density of the commu-
nities compared to their expected density if the links were randomly distributed
over the network.
In order to evaluate Louvain’s performance, and for fair comparison, we first
project the bipartite graph over the > nodes, generating a weighted graph ac-
cording to the similarity function θ (common neighbor in our case). Then we
apply Louvain on the weighted graph.
Infomap: Infomap is a recursive algorithm, similar to Louvain, where each node
is moved to a neighboring community if this modification minimizes the length
of the map equation [19]. Infomap can account for the bipartite structure and we
use this feature to generate a partition of > nodes only.
LP BRIM: LP BRIM [10] is a community detection algorithm that optimizes the
bimodularity [20] which is an extension of the modularity for bipartite graphs. It
relies on BRIM algorithm (Bipartite, Recursively Induced Modules) and uses a
label propagation procedure.
Because LP BRIM provides a partition of the complete bipartite networks – com-
munities are composed of > and ⊥ nodes –, we adapt the algorithm and define
a community by keeping only the > nodes of the partitions. This allows a fair
comparison in the evaluation process.

3 Evaluation of C OM S IM

This section is devoted to assess the relevance of the proposed method. We start by
investigating how the different algorithms behave on two small networks equipped
with existing communities (Section 3.1) before showing how C OM S IM scales up
when dealing with large-scale networks (Section 3.2).

3.1 On dataset with ground-truth communities

We first apply our algorithm to two networks which are small but are provided with
a notion of ground-truth communities that we use as a reference to compare the four
algorithms.

Southern women [21] is a network depicting the participation of 18 women to 14

events in the United States observed during a nine-months period in 1930. Al-
though small, this dataset is very interesting since it has been extensively studied
by social scientists to understand how social groups form and evolve (see [22, 23]
for instance). In this study, we use the partition found in the literature as the
ground-truth communities to which we compare the four algorithms.
C OM S IM: A bipartite community detection algorithm using cycle and node’s similarity 7

(a) NMI (b) F1-score

Fig. 3 Evaluation of the quality of the partitions detected by the algorithms on 20 newsgroups and
Southern Women.

20 newsgroups [24] is a record of approximately 50 000 posts submitted by 30 000

users (bot) over 20 groups of discussion (>).
Figure 3 presents the results of the comparison between C OM S IM and the three
baseline algorithms for bipartite community detection described in Section 2.3. All
algorithms were applied on Southern Women and 20 newsgroups dataset 100 times.
The box plots provides the maximal, minimal and average values.
Since we have a ground-truth partition for the dataset, we use first the usual
Normalized Mutual Information (NMI, see [25] for instance) to compute how far
the detected partitions are from the ground-truth ones. Figure 3a reveals that for
both datasets C OM S IM is the algorithm that proposes the best partition in average.
One can also notice that Infomap and LP BRIM generate good partitions for 20
newsgroups and Louvain good partitions for Southern Women.
It is interesting to notice that Infomap completely fails to detect the expected
communities for Southern Women. Manual investigation revealed that all women of
Southern Women are actually gathered in a single community, which is well captured
by NMI (NMI= 0). On the opposite, each node of 20 newsgroups is positioned
in a different community, which is completely overestimated by the NMI (NMI=
0.7643).
In order to provide a second point of view, we also use the F1-score, a classical
metric to evaluate the performance of prediction algorithms (see [26] for an example
of F1-score used in the context of community detection issues). Figure 3b shows
again that C OM S IM is the best community detection algorithm for both datasets in
average. Interestingly, for this metric, Louvain seems to propose good partitions in
average and for both dataset.
All in all, it seems that C OM S IM proposes coherent communities when com-
pared to ground-truth partitions of bipartite networks. The next section intends to
investigate how the algorithm behaves on a large scale network.
8 Raphael Tackx, Fabien Tarissan, and Jean-Loup Guillaume

3.2 On large-scale networks

In order to test the performance of our algorithm both in terms of efficiency and
quality, we rely here on a large dataset extracted from the Internet Movie Database
(IMDb). This dataset [27] presents a bipartite network composed of 118 258 actors
(⊥) who played in 122 131 movies (>) between 1980 and 20103 .

Southern women 20 newsgroups IMDb

|>|/|⊥|/links 18/14/89 20/30K/42K 122K/118K/531K
C OM S IM 1.7 ms / 11.5 MB 1.1 s / 30 MB 33.5 s / 591.6 MB
Infomap 13 ms / 10.7 MB 951 ms / 6.3 MB 100s / 374 MB
Louvain 11 ms / 6.5 MB 86 ms / 10.1 MB 21 s / 43 MB
LP BRIM 6.7 s / 6.1 MB 74.2 s / 59.7MB -/-
Table 1 Performances in terms of execution time and memory peak for the four algorithms.

Table 1 presents the performances in terms of execution time and memory peak
for the four algorithms on the three datasets. This shows that Louvain remains the
most efficient algorithm in terms both of time and memory, revealing to be slower
only on the smallest dataset.
However, it should be highlighted here that the performances of Louvain shown
in Table 1 have been recorded after the >-projection. This means that part of the
computation load related to the θ function has been avoided, which is not the case
for the other algorithms. It thus mechanically favour the Louvain approach.
To that regard, it is worth noticing that our algorithm presents good results. On
IMDb in particular, C OM S IM is only slightly slower than Louvain and three times
faster than Infomap.

The results above show that our algorithm can scale up to large networks but that
it provides no insight on the quality of the detected communities. In contrast to the
previous section where we had ground-truth knowledge of the good partitions, no
study conducted on the IMDb dataset proposes an objective and external partition
of the nodes. It is thus impossible to use here either NMI or F1-score to compare
the three remaining algorithms 4 .
In order to assess the quality of the proposed communities, we follow instead
the proposition made in [28] where the authors introduce two goodness functions in
an attempt to quantify how relevant a community is regarding two properties that
we adapted for the case of bipartite graphs: the Density (or Internal Density) and
Separability.

3 For an homogeneous analysis, we removed all TV shows and documentaries and kept only the 7
first actors listed in the casting.
4 Since LP BRIM does not scale up to the size of IMDb, we avoid mentioning this approach in the

rest of the study.

C OM S IM: A bipartite community detection algorithm using cycle and node’s similarity 9

(a) Internal Density (b) Separability

Fig. 4 Scatter plot displaying the relation between properties of the communities and their size for
C OM S IM (top), Louvain (middle) and Infomap (bottom) on IMDb.

More formally, let P be a partition of > nodes. Given a community Ci ∈ P, we

N(xi ) the set of ⊥ nodes induced by the neighborhood of Ci . Let
S
denote by Ci =
xi ∈Ci
mCi be the number of edges between Ci and Ci (internal number of edges) and mCi
the number of edges between Ci and C j (external number of edges), where j 6= i.
Then we define the internal density and the separability as follow:
mCi
• Internal Density of community Ci : |Ci |∗|Ci |
mCi
• Separability of community Ci : mCi +mC
i

Those two indicators allow to evaluate how coherent a community is regarding

internal and external edges. Figure 4 presents the distribution of the Internal Density
(Figure 4a) and the Separability (Figure 4b) of communities according to their size
and for the three algorithms (C OM S IM top, Louvain middle and Infomap bottom).
This shows that Infomap mostly fails to detect coherent communities. Indeed,
although most of the communities have very high values for both properties, it con-
cerns mostly very little communities composed of few nodes. But for large commu-
nities, the indicators drop. This is particularly obvious on Figure 4a. To that regard,
it is striking to notice that the largest community detected by Infomap gathers more
than 46% of the nodes and the 7 largest communities involve more than 96% of the
nodes. The same observation can be made for Louvain, although to a lesser extent.
The largest community involves 29% (48% for the 7 largest communities).
Compared to Louvain and Infomap, C OM S IM proposes more balanced commu-
nities in terms of size. The largest community is rather small (only 1% of the nodes)
while keeping a profile of density close to the ones of Louvain (see Figure 4a). Re-
garding the separability however, there is a slight shift towards low values compared
to Louvain, which indicates that the quality of the partitions could be improved for
this property.
All in all, an although more experiments should be made in order to complete
the comparison, we claim that this study is a first step establishing the relevance
10 Raphael Tackx, Fabien Tarissan, and Jean-Loup Guillaume

of the proposed approach, both in terms of efficiency and quality of the detected
communities.

4 Conclusions

In this study we proposed C OM S IM, a new algorithm to detect community in bi-

partite networks. This approach generates a partition of the > nodes by relying on
similarity between the nodes in terms of connections towards ⊥ nodes. To do so,
it tries to find and maximize cycles of relations between > nodes. This defines the
core of the communities which are enriched with new nodes during a second phase
of the algorithm.
We implemented and applied this algorithm on 3 datasets and compared the gen-
erated partitions with the ones proposed by three baseline algorithms used on bi-
partite graphs. The empirical results showed that, on small networks for which we
had a ground-truth knowledge of the good partition, C OM S IM is the algorithm that
generates the best communities.
In addition, C OM S IM proved to scale up with a time complexity close to Louvain.
Investigating qualitatively the partitions, we showed that the communities generated
by C OM S IM are more balanced in terms of size, while keeping quality indicators
reasonable and comparable to the ones proposed by Louvain for instance.
It is worth noticing that other algorithms could have been used for the compar-
ison. For instance, biSBM [11] or SCD [12] are relevant, although not completely
adapted to this context. The former requires to provide the number of expected com-
munities, while the latter proposes overlapping communities.
We claim that this study establishes the relevance of the approach and we let
more in-depth study for future work.

Acknowledgements

This work is funded in part by the European Commission H2020 FETPROACT

2016-2017 program under grant 732942 (ODYCCEUS), by the ANR (French Na-
tional Agency of Research) under grants ANR-15-CE38-0001 (AlgoDiv) and ANR-
13-CORD-0017-01 (CODDDE), by the French program ”PIA - Usages, services et
contenus innovants” under grant O18062-44430 (REQUEST), and by the Ile-de-
France program FUI21 under grant 16010629 (iTRAC).
C OM S IM: A bipartite community detection algorithm using cycle and node’s similarity 11

References

1. Duncan J Watts and Steven H Strogatz. Collective dynamics of ’small-world’ networks. na-
ture, 393(6684):440–442, 1998.
2. Ramon Ferrer i Cancho and Richard V Solé. The small world of human language. Proceedings
of the Royal Society of London. Series B: Biological Sciences, 268(1482):2261–2265, 2001.
3. Mark EJ Newman, Duncan J Watts, and Steven H Strogatz. Random graph models of social
networks. Proceedings of the National Academy of Sciences of the United States of America,
99(Suppl 1):2566–2572, 2002.
4. Stefano Battiston and Michele Catanzaro. Statistical properties of corporate board and direc-
tor networks. The European Physical Journal B-Condensed Matter and Complex Systems,
38(2):345–352, 2004.
5. Fabrice Le Fessant, Sidath Handurukande, A-M Kermarrec, and Laurent Massoulié. Clus-
tering in peer-to-peer file sharing workloads. In Peer-to-Peer Systems III, pages 217–226.
Springer, 2005.
6. Christophe Prieur, Dominique Cardon, Jean-Samuel Beuscart, Nicolas Pissard, and Pas-
cal Pons. The stength of weak cooperation: A case study on flickr. arXiv preprint
arXiv:0802.2317, 2008.
7. Yong-Yeol Ahn, Sebastian E Ahnert, James P Bagrow, and Albert-László Barabási. Flavor
network and the principles of food pairing. Scientific reports, 1, 2011.
8. Santo Fortunato. Community detection in graphs. Physics reports, 486(3):75–174, 2010.
9. Mark EJ Newman, Steven H Strogatz, and Duncan J Watts. Random graphs with arbitrary
degree distributions. Physics Reviews E, 64, 2001.
10. Xin Liu and Tsuyoshi Murata. Community detection in large-scale bipartite networks. In
Proceedings of the 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence
and Intelligent Agent Technology - Volume 01, WI-IAT ’09, pages 50–57, Washington, DC,
USA, 2009. IEEE Computer Society.
11. Daniel B Larremore, Aaron Clauset, and Abigail Z Jacobs. Efficiently inferring community
structure in bipartite networks. Physical Review E, 90(1):012805, 2014.
12. Arnau Prat-Pérez, David Dominguez-Sal, and Josep-Lluis Larriba-Pey. High quality, scalable
and parallel community detection for large real graphs. In Proceedings of the 23rd interna-
tional conference on World wide web, pages 225–236. ACM, 2014.
13. Sune Lehmann, Martin Schwartz, and Lars Kai Hansen. Biclique communities. Physical
Review E, 78(1):016108, 2008.
14. Paul Jaccard. Le coefficient generique et le coefficient de communaute dans la flore marocaine.
Impr. Commerciale, 1926.
15. Tao Zhou, Linyuan Lü, and Yi-Cheng Zhang. Predicting missing links via local information.
The European Physical Journal B-Condensed Matter and Complex Systems, 71(4):623–630,
2009.
16. Lada A Adamic and Eytan Adar. Friends and neighbors on the web. Social networks,
25(3):211–230, 2003.
17. Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre. Fast
unfolding of communities in large networks. Journal of statistical mechanics: theory and
experiment, 2008(10):P10008, 2008.
18. Mark EJ Newman. Modularity and community structure in networks. Proceedings of the
national academy of sciences, 103(23):8577–8582, 2006.
19. Martin Rosvall, Daniel Axelsson, and Carl T Bergstrom. The map equation. The European
Physical Journal-Special Topics, 178(1):13–23, 2009.
20. M. J. Barber. Modularity and community detection in bipartite networks. Physical Review E,
76(6):066102, December 2007.
21. Allison Davis, Burleigh B. Gardner, and Mary R. Gardner. Deep South; a Social Anthropo-
logical Study of Caste and Class. The University of Chicago Press, Chicago, 1941.
22. Elna C Green. Southern strategies: Southern women and the woman suffrage question. Univ
of North Carolina Press, 1997.
12 Raphael Tackx, Fabien Tarissan, and Jean-Loup Guillaume

23. Linton C Freeman. Finding social groups: A meta-analysis of the southern women data. na,
2003.
24. Ken Lang. Newsweeder: Learning to filter netnews. In Proceedings of the Twelfth Interna-
tional Conference on Machine Learning, pages 331–339, 1995.
25. Andrea Lancichinetti, Santo Fortunato, and János Kertész. Detecting the overlapping and hi-
erarchical community structure in complex networks. New Journal of Physics, 11(3):033015,
2009.
26. Jaewon Yang and Jure Leskovec. Overlapping community detection at scale: a nonnegative
matrix factorization approach. In Proceedings of the sixth ACM international conference on
Web search and data mining, pages 587–596. ACM, 2013.
27. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christo-
pher Potts. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual
Meeting of the Association for Computational Linguistics: Human Language Technologies,
pages 142–150, Portland, Oregon, USA, June 2011. Association for Computational Linguis-
tics.
28. Jaewon Yang and Jure Leskovec. Defining and evaluating network communities based on
ground-truth. Knowledge and Information Systems, 42(1):181–213, 2015.