ArticlePDF Available

Abstract and Figures

We propose a simple method to extract the community structure of large networks. Our method is a heuristic method that is based on modularity optimization. It is shown to outperform all other known community detection methods in terms of computation time. Moreover, the quality of the communities detected is very good, as measured by the so-called modularity. This is shown first by identifying language communities in a Belgian mobile phone network of 2 million customers and by analysing a web graph of 118 million nodes and more than one billion links. The accuracy of our algorithm is also verified on ad hoc modular networks.
Content may be subject to copyright.
Fast unfolding of communities in large networks
This article has been downloaded from IOPscience. Please scroll down to see the full text article.
J. Stat. Mech. (2008) P10008
Download details:
IP Address:
The article was downloaded on 30/12/2012 at 07:21
Please note that terms and conditions apply.
View the table of contents for this issue, or go to the journal homepage for more
Home Search Collections Journals About Contact us My IOPscience
J. Stat. Mech.
(2008) P10008
ournal of Statistical Mechanics:
An IOP and SISSA journal
Theory and Experiment
Fast unfolding of communities in large
Vincent D Blondel
, Jean-Loup Guillaume
Renaud Lambiotte
and Etienne Lefebvre
Departmen t of Mathematical Engineering, Universit´e Catholique de Louvain,
4 avenue Georges Lemaitre, B-1348 Louvain-la-Neuve, Belgium
LIP6, Universit´e Pierre et Marie Curie, 4 place Jussieu, F-75005 Paris, France
Institute for Mathematical Sciences, Imperial College London,
53 Prince’s Gate, South Kensington Campus, London SW7 2PG, UK
E-mail:,, and
Receiv ed 18 April 2008
Accepted 3 September 2008
Published 9 October 2008
Online at
Abstract. We propose a simple method to extract the community structure of
large networks. Our method is a heuristic method that is based on modularity
optimization. It is shown to outperform all other known community detection
methods in terms of computation time. Moreover, the quality of the communities
detected is very good, as measured by the so-called modularit y. This is shown
first by identifying language communities in a Belgian mobile phone network of
2 million customers and b y analysing a web graph of 118 million nodes and more
than one billion links. The accuracy of our algorithm is also verified on ad hoc
mo dular networks.
Keywords: random graphs, networks, new applications of statistical mechanics
ArXiv ePrint: 0803.0476
2008 IOP Publishing Ltd and SISSA 1742-5468/08/P10008+12$30.00
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
1. Introduction 2
2. Method 3
3. Application to large networks 6
4. Conclusion and discussion 10
Acknowledgments 11
References 11
1. Introduction
Social, technological and information systems can often be described in terms of complex
networks that have a topology of interconnected nodes combining organization and
randomness [1, 2]. The typical size of large networks such as social network services,
mobile phone networks or the web is now counted in millions, if not billions, of nodes
and these scales demand new methods to retrieve comprehensive information from their
structure. A promising approach consists in decomposing the networks into sub-units
or communities, which are sets of highly interconnected nodes [3]. The identification of
these communities is of crucial importance as they may help to uncover aprioriunknown
functional modules such as topics in information networks or cyber-communities in social
networks. Moreover, the resulting meta-network, whose nodes are the communities, may
then be used to visualize the original network structure.
The problem of community detection requires the partition of a network into
communities of densely connected nodes, with the nodes belonging to different
communities being only sparsely connected. Precise formulations of this optimization
problem are known to be computationally intractable. Several algorithms have therefore
been proposed to find reasonably good partitions in a reasonably fast way. This search
for fast algorithms has attracted much interest in recent years due to the increasing
availability of large network datasets and the impact of networks on everyday life. One
can distinguish several types of community detection algorithms: divisive algorithms
detect inter-community links and remove them from the network [4]–[6], agglomerative
algorithms merge similar nodes/communities recursively [7] and optimization methods
are based on the maximization of an objective function [8]–[10]. The quality of the
partitions resulting from these methods is oftenmeasuredbytheso-calledmodularity
of the partition. The modularity of a partition is a scalar value between 1and1
that measures the density of links inside communities as compared to links between
communities [5, 11]. In the case of weighted networks (weighted networks are networks
that have weights on their links, such as the number of communications between two
mobile phone users), it is defined as [12]
Q =
), (1)
doi:10.1088/1742-5468/2008/10/P10008 2
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
where A
represents the weight of the edge between i and j, k
is the sum of
the weights of the edges attached to vertex i, c
is the community to which vertex i is
assigned, the δ function δ(u, v)is1ifu = v and 0 otherwise and m =
Modularity has been used to compare the quality of the partitions obtained by
different methods, but also as an objective function to optimize [13]. Unfortunately,
exact modularity optimization is a problem that is computationally hard [14]andso
approximation algorithms are necessary when dealing with large networks. The fastest
approximation algorithm for optimizing modularity on large networks was proposed by
Clauset et al [8]. That method consists in recurrently merging communities that optimize
the production of modularity. Unfortunately, this greedy algorithm may produce values
of modularity that are significantly lower than what can be found by using, for instance,
simulated annealing [15]. Moreover, the method proposed in [8] has a tendency to produce
super-communities that contain a large fraction of the nodes, even on synthetic networks
that have no significant community structure. This artefact also has the disadvantage
of slowing down the algorithm considerably and makes it inapplicable to networks of
more than a million nodes. This undesired effect has been circumvented by introducing
tricks in order to balance the size of the communities being merged, thereby speeding up
the running time and making it possible to deal with networks that have a few million
nodes [16].
The largest networks that have been dealt with so far in the literature are a protein–
protein interaction network of 30 739 nodes [17], a network of about 400 000 items on sale
on the website of a large on-line retailer [8] and a Japanese social networking system of
about 5.5 million users [16]. These sizes still leave considerable room for improvement [18]
considering that, as of today, the social networking service Facebook has about 64 million
active users, the mobile network operator Vodafone has about 200 million customers and
Google indexes several billion web-pages. Let us also notice that in most large networks
such as those listed above there are several natural organization levels—communities
divide themselves into sub-communities—and it is thus desirable to obtain community
detection methods that reveal this hierarchical structure [19].
2. Method
We now introduce our algorithm that finds high modularity partitions of large networks
in a short time and that unfolds a complete hierarchical community structure for the
network, thereby giving access to different resolutions of community detection. Contrary
to all the other community detection algorithms, the network size limits that we are facing
with our algorithm are due to limited storage capacity rather than limited computation
time: identifying communities in a 118 million nodes network took only 152 min
Our algorithm is divided into two phases that are repeated iteratively. Assume that
we start with a weighted network of N nodes. First, we assign a different community to
each node of the network. So, in this initial partition there are as many communities as
there are nodes. Then, for each node i we consider the neighbours j of i and we evaluate
the gain of modularity that would take place by removing i from its community and by
All methods described here have been compiled and tested on the same machine: a bi-opteron 2.2k with 24 GB
of memory. The code is freely available for download on the web-page
doi:10.1088/1742-5468/2008/10/P10008 3
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
placing it in the community of j. The node i is then placed in the community for which
this gain is maximum (in the case of a tie we use a breaking rule), but only if this gain
is positive. If no positive gain is possible, i stays in its original community. This process
is applied repeatedly and sequentially for all nodes until no further improvement can be
achieved and the first phase is then complete. Let us insist on the fact that a node may be,
and often is, considered several times. This first phase stops when a local maxima of the
modularity is attained, i.e. when no individual move can improve the modularity
should also note that the output of the algorithm depends on the order in which the nodes
are considered. Preliminary results on several test cases seem to indicate that the ordering
of the nodes does not have a significant influence on the modularity that is obtained, while
it may affect the computation time, but the reasons for this dependence are not clear. In
particular, taking the nodes in a natural order related to the community structure itself
(e.g. the order given by a previous community computation, or the postcode) does not
give clear improvement (see section 3). The problem of choosing an order is thus worth
studying since it could give good heuristics to enhance the computation time.
Part of the algorithm’s efficiency results from the fact that the gain in modularity
ΔQ obtained by moving an isolated node i into a community C can easily be computed
ΔQ =
, (2)
is the sum of the weights of the links inside C,
is the sum of the weights
of the links incident to nodes in C, k
is the sum of the weights of the links incident to
node i, k
is the sum of the weights of the links from i to nodes in C and m is the sum
of the weights of all the links in the network. A similar expression is used in order to
evaluate the change of modularity when i is removed from its community. In practice,
one therefore evaluates the change of modularity by removing i from its community and
then by moving it into a neighbouring community.
The second phase of the algorithm consists in building a new network whose nodes
are now the communities found during the first phase. To do so, the weights of the links
between the new nodes are given by the sum of the weight of the links between nodes
in the corresponding two communities [20]. Links between nodes of the same community
lead to self-loops for this community in the new network. Once this second phase is
completed, it is then possible to reapply the first phase of the algorithm to the resulting
weighted network and to iterate. Let us denote by ‘pass’ a combination of these two
phases. By construction, the number of meta-communities decreases at each pass, and
as a consequence most of the computing time is used in the first pass. The passes are
iterated (see figure 1) until there are no more changes and a maximum of modularity is
attained. The algorithm is reminiscent of the self-similar nature of complex networks [21]
and naturally incorporates a notion of hierarchy, as communities of communities are built
during the process. The height of the hierarchy that is constructed is determined by the
number of passes and is generally a small number, as will be shown in some examples
In order to decrease the overall running time of the method it is possible to introduce a threshold and then stop
the first phase as soon as the relative gain in modularity does not exceed this threshold. The numerical results
reported here have been obtained with this minor modification.
doi:10.1088/1742-5468/2008/10/P10008 4
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
Figure 1. Visualization of the steps of our algorithm. Each pass is made of
two phases: one where modularity is optimized by allowing only local changes
of comm unities; one where the communities found are aggregated in order to
build a new network of communities. The passes are repeated iteratively until
no increase of mo dularity is possible.
This simple algorithm has several advantages. First, its steps are intuitive and easy to
implement, and the outcome is unsupervised. Moreover, the algorithm is extremely fast,
i.e. computer simulations on large ad hoc modular networks suggest that its complexity
is linear on typical and sparse data. This is due to the fact that the possible gains
in modularity are easy to compute with the above formula and that the number of
communities decreases drastically after just a few passes so that most of the running
time is concentrated on the first iterations. The so-called resolution limit problem of
modularity also seems to be circumvented thanks to the intrinsic multi-level nature of our
algorithm. Indeed, it is well known [22] that modularity optimization fails to identify
communities smaller than a certain scale, thereby inducing a resolution limit on the
community detected by a pure modularity optimization approach. This observation is
only partially relevant in our case because the first phase of our algorithm involves
the displacement of single nodes from one community to another. Consequently, the
probability that two distinct communities can be merged by moving nodes one by one is
very low. These communities may possibly be merged in the later passes, after blocks
of nodes have been aggregated. However, our algorithm provides a decomposition of the
network into communities for different levels of organization. In order to illustrate this
feature, let us focus on the ring of 30 cliques discussed in [22], where the cliques are
composed of 5 nodes and are interconnected through single links. The first pass of the
algorithm finds the natural partition of the network, where each community corresponds
to one clique. The second pass finds the global maximum of modularity where cliques
are combined into groups of 2. Consequently, if the cliques are indeed merged in the final
doi:10.1088/1742-5468/2008/10/P10008 5
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
Table 1. Summary of numerical results. This table gives the performances of the
algorithm of Clauset et al [8], of Pons and Latap y [7], of Wakita and Tsurumi [16]
and of our algorithm for community detection in networks of various sizes. For
each method/network, the table displays the modularity that is achieved and the
computation time. Empty cells correspond to a computation time over 24 h. Our
method clearly performs better in terms of computer time and modularity. It is
also interesting to note the small value of Q found by WT for the mobile phone
network. This bad modularity result may originate from their heuristic which
creates balanced comm unities, while our approach gives unbalanced communities
in this specific network.
Karate Arxiv Internet Web Phone Web uk-2005
Nodes/ 34/77 9k/24k 70k/351k 325k/1M 2.04M/5.4M 39M/783M 118M/1B
CNM 0.38/0 s 0.772/3.6 s 0.692/799 s 0.927/5034 s —/— —/— —/—
PL 0.42/0 s 0.757/3.3 s 0.729/575 s 0.895/6666 s —/— —/— —/—
WT 0.42/0 s 0.761/0.7 s 0.667/62 s 0.898/248 s 0.553/367 s —/— —/—
Our 0.42/0 s 0.813/0 s 0.781/1 s 0.935/3 s 0.76/44 s 0.979/738 s 0.984/152 mn
partition due to the resolution limit, they are distinct after the first pass. This result
suggests that the intermediate solutions found by our algorithm may also be meaningful
and that the uncovered hierarchical structure may allow the end-user to zoom into the
network and to observe its structure with the desired resolution.
3. Application to la rge net wo rks
In order to verify the validity of our algorithm, we have applied it on a number of test-
case networks that are commonly used for efficiency comparison and we have compared
it with three other community detection algorithms (see table 1). The networks that
we consider include a small social network [23], a network of 9000 scientific papers and
their citations [24], a sub-network of the internet [25] and a web-page network of a few
hundred thousand web-pages (the domain, see [26]). In all cases, one can observe
both the rapidity and the large values of the modularity that are obtained. Our method
outperforms all the other methods to which it is compared. We also have applied our
method on two web networks of unprecedented sizes: a sub-network of domain
of 39 million nodes and 783 million links [27] and a network of 118 million nodes and 1
billion links obtained by the Stanford WebBase crawler [27, 28]. Even for these very large
networks, the computation time is small (12 min and 152 min, respectively) and makes
networks of still larger size, perhaps a billion nodes, accessible to computational analysis.
It is also interesting to note that the number of passes is usually very small. In the case
of the Karate Club [23], for instance, there are only 3 passes: during the first one, the 34
nodes of the network are partitioned into 6 communities; after the second one, only four
communities remain; during the third one, nothing happens and the algorithm therefore
stops. In the above examples, the number of passes is always smaller than 5.
doi:10.1088/1742-5468/2008/10/P10008 6
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
We have also tested the sensitivity of our algorithm by applying it on ad hoc networks
that have a known community structure. To do so, we have used networks composed of
128 nodes which are split into 4 communities of 32 nodes each [29]. Pairs of nodes
belonging to the same community are linked with probability p
while pairs belonging
to different communities are linked with probability p
. The accuracy of the method
is evaluated by measuring the fraction of correctly identified nodes and the normalized
mutual information. In the benchmark proposed in [29], the fraction of correctly identified
nodes is 0.67 for z
=8,0.92 for z
=7and0.98 for z
= 6, i.e. an accuracy similar
to that of the algorithm of Pons and Latapy [7] and of the algorithm of Reichardt and
Bornholdt [30]. To our knowledge, only two algorithms have a better accuracy than ours,
the algorithm of Duch and Arenas [31] and the simulated annealing method first proposed
in [15], but their computational cost limits their applicability to much smaller networks
than the ones considered here. Our algorithm has also been successfully tested on other
benchmarks, such as the ones proposed in [19, 32]. In the case [32], the normalized mutual
information is nearly 1 for the macro-communities with a mixing parameter k
up to 35.
It reaches 0.5 when the mixing parameter is around 55.
To validate the communities obtained we have also applied our algorithm to a large
network constructed from the records of a Belgian mobile phone company. This network is
describedindetailin[33] where it is shown to exhibits typical features of social networks,
such as a high clustering coefficient and a fat-tailed degree distribution. The network is
composed of 2.6 million customers, between whom weighted links are drawn that account
for their total number of phone calls during a 6 month period. In this paper, we have
focused on a subset of 2.04 million customers for whom several entries are associated,
such as their age, sex, language and the postcode of the place where they live. This
large social network is exceptional due to the particular situation of Belgium where two
main linguistic communities (French and Dutch) coexist and which provides an excellent
way to test the validity of our community detection method by looking at the linguistic
homogeneity of communities [34]. From a more sociological point of view, the possibility to
highlight the linguistic, religious or ethnic homogeneity of communities opens perspectives
for describing the social cohesion and the potential fragility of a country [35].
On this particular network, our community detection algorithm has identified a
hierarchy of 6 levels. At the bottom level every customer is a community of its own
and at the top level there are 261 communities that have more than 100 customers. These
communities account for about 75% of all customers. We have performed a language
analysis of these 261 communities (see figure 2). The homogeneity of a community
is characterized by the percentage of those speaking the dominant language in that
community; this quantity goes to 1 when the community tends to be monolingual. Our
analysis reveals that the network is strongly segregated, with most communities almost
monolingual. There are 36 communities with more than 10 000 customers and, except for
one community at the interface between the two language clusters, all these communities
have more than 85% of their members speaking the same language (see figure 3 for a
complete distribution). It is interesting to analyse more closely the only community
that has a more equilibrate distribution of languages. Our hierarchy-revealing algorithm
allows us to do this by considering the sub-communities provided by the algorithm at
the lower level. As shown in figure 3, these sub-communities are closely connected
to each other and are themselves composed of heterogeneous groups of people. These
doi:10.1088/1742-5468/2008/10/P10008 7
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
Figure 2. Graphical representation of the network of communities extracted from
a Belgian mobile phone network. About 2 million customers are represented on
this network. The size of a node is proportional to the number of individuals in the
corresponding communit y and its colour on a red–green scale represents the main
language spoken in the community (red for French and green for Dutch). Only the
communities composed of more than 100 customers have been plotted. Notice
the intermediate communit y of mixed colours between the two main language
clusters. A zoom at higher resolution reveals that it is made of sev eral sub-
communities with less apparent language separation.
groups of people, where language ceases to be a discriminating factor, might possibly
play a crucial role for the integration of the country and for the emergence of consensus
between the communities [36]. One may indeed wonder what would happen if the
community at the interface between the two language clusters in figure 2 was to be
doi:10.1088/1742-5468/2008/10/P10008 8
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
Figure 3. For the largest communities in the Belgian mobile phone network
w e represent the size of the community and the proportion of customers in the
community that sp eak the dominant language of the community. For all but one
community of more than 10 000 members the dominant language is spoken by
more than 85% of the community members.
Another interesting observation is related to the presence of other languages. There
are actually four possible language declarations for the customers of this particular mobile
phone operator: French, Dutch, English or German. It is interesting to note that, whereas
English-speaking customers disperse themselves quite evenly in all communities, more
than 60% of the German speaking customers are concentrated in just one community.
This is probably due to the fact that German-speaking people are mainly concentrated in
a small region close to Germany, while English-speaking people are spread over the whole
country. Let us finally observe that, as can be visually noticed in figure 2, French-speaking
communities are much more densely connected than their Dutch-speaking counterparts:
on average, the strength of the links between French-speaking communities is 54%
stronger than those between Dutch-speaking communities. This difference of structure
between the two sub-networks seems to indicate that the two linguistic communities
are characterized by different social behaviours and therefore suggests to search other
topological characteristics for the communities.
We have also focused on this mobile phone network in order to elucidate the role
played by the ordering of the nodes on the output of the method. To do so, we have
first performed 100 analyses where the ordering of the nodes is chosen randomly. From
the modularity Q
and computation time T
evaluated at each run i,wehavecomputed
the average and variance of these variables. In the case of modularity, the value is almost
constant over all the runs with Q =0.76 and a deviation of σ
The fluctuations of the computation time are more important, as T =44.2sand
doi:10.1088/1742-5468/2008/10/P10008 9
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
=3.2 s but remain reasonably small as this is only a 7% variation. The smallest
and largest values of T among the 100 runs are 39 and 55 s. This interval suggests that a
good choice for the ordering of the nodes may substantially accelerate the dynamics. We
have therefore checked if an order related to the community structure would accelerate
the computation time. To do so, we have ordered the nodes by their postcodes, but this
choice did not lead to any improvement as compared to a random ordering, as Q
and T
4. Conclusion and discussion
We have introduced an algorithm for optimizing modularity that allows us to study
networks of unprecedented size. The limitation of the method for the experiments that we
performed was the storage of the network in main memory rather than the computation
time. This change of scales, i.e. from around 5 millions nodes for previous methods
to more than 100 million nodes in our case, opens exciting perspectives as the modular
structure of complex systems such as whole countries or huge parts of the Internet can
now be unravelled. The accuracy of our method has also been tested on ad hoc modular
networks and is shown to be excellent in comparison with other (much slower) community
detection methods. It is interesting to note that the speed of our algorithm can still be
substantially improved by using some simple heuristics, for instance by stopping the first
phase of our algorithm when the gain of modularity is below a given threshold or by
removing the nodes of degree 1 (leaves) from the original network and adding them back
after the community computation. The impact of these heuristics on the final partition
of the network should be studied further, as well as the role played by the ordering of the
nodes during the first phase of the algorithm.
By construction, our algorithm unfolds a complete hierarchical community structure
for the network, each level of the hierarchy being given by the intermediate partitions found
at each pass. In this paper, however, we have only verified the accuracy of the top level of
this hierarchy, namely the final partition found by our algorithm, and the accuracy of the
intermediate partitions has still to be shown. Several points suggest, however, that these
intermediate partitions make sense. First, intermediate partitions correspond to local
maxima of modularity, maxima in the sense that it is not possible to increase modularity
by moving one single ‘entity’ from one community to a neighbouring one. In the first pass
of the algorithm, these entities are nodes, but at subsequent passes, they correspond to
larger and larger sets of nodes. Intermediate partitions may therefore be viewed as local
maxima of modularity at different scales. It is the agglomeration of nodes during the
second phase of the algorithm which allows us to uncover larger and larger communities,
thereby taking advantage of the self-similar structure of many complex networks. Second,
the final partition found by our algorithm has a very high value of modularity for a broad
range of system sizes (for instance, as shown in table 1, our algorithm performs better
in terms of modularity than those of Clauset et al [8], of Pons and Latapy [7]andof
Wakita and Tsurumi [16]). Finally, it is instructive to consider a community C found at
the last pass of our algorithm. In order to test the validity of the sub-communities found
at the penultimate pass, it is tempting to look at community C as a new network, thereby
neglecting links going from C to the rest of the network. By reapplying our algorithm on
the isolated community C, one expects to find very similar sub-communities due to the
doi:10.1088/1742-5468/2008/10/P10008 10
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
local optimization involved at each step. These are, however, very qualitative arguments
and the multi-resolution of our algorithm will only be confirmed after looking in detail at
the hierarchies found in ad hoc networks with known hierarchical structure [19] or without
community structure (e.g. Erd¨os–Renyi random graphs), or after comparing with other
methods incorporating a tunable resolution [32, 37, 38].
This research was supported by the Communaut´eFran¸caise de Belgique through a grant
ARC and by the Belgian Network DYSCO, funded by the Interuniversity Attraction Poles
Programme, initiated by the Belgian State, Science Policy Office. J-LG is also supported
in part by MAPAP SIP-2006-PP-221003 and ANR MAPE projects.
[1] Albert R and Barab´asi A-L, 2002 Rev. Mod. Phys. 74 4797
[2] Newman M E J, Barab´asi A-L and Watts D J, 2006 The Structure and Dynamics of Networks (Princeton,
NJ: Princeton University Press)
[3] Fortunato S and Castellano C, 2007 arXiv:0712.2716
[4] Girvan M and Newman M E J, 2002 Proc. Nat. A cad. Sci. 99 7821
[5] Newman M E J and Girvan M, 2004 Phys. Rev. E 69 026113
[6] Radicchi F, Castellano C, Cecconi F, Loreto V and Parisi D, 2004 Proc. Nat. Acad. Sci. 101 2658
[7] Pons P and Latapy M, 2006 J. Graph Algorithms Appl. 10 191
[8] Clauset A, Newman M E J and Moore C, 2004 Phys. Rev. E 70 066111
[9] Wu F and Huberman B A, 2004 Eur. Phys. J. B 38 331
[10] Newman M E J, 2006 Phys. Rev. E 74 036104
[11] Newman M E J, 2006 Proc. Nat. Acad. Sci. 103 8577
[12] Newman M E J, 2004 Phys. Rev. E 70 056131
[13] Newman M E J, 2004 Phys. Rev. E 69 066133
[14] Brandes U, Delling D, Gaertler M, Goerke R, Hoefer M, Nikoloski Z and Wagner D, 2006
[15] Guimera R, Sales M and Amaral L A N, 2004 Phys. Rev. E 70 025101
[16] Wakita K and Tsurumi T, 2007 Proc. IADIS Int. Conf. on WWW/Internet 2007 p 153
[17] Palla G, Der´enyi I, Farkas I and Vicsek T, 2005 Nature 435 814
[18] Raghavan U N, Albert R and Kumara S, 2007 Phys. Rev. E 76 036106
[19] Sales-Pardo M, Guimera R, Moreira A A and Amaral L A N, 2007 Proc. Nat. Acad. Sci. 104 15224
[20] Arenas A, Duch J, Fern´andez A and G´omez S, 2007 New J. Phys. 9 176
[21] Song C, Havlin S and Makse H A, 2005 Nature 433 392
[22] Fortunato S and Barth´elemy M, 2007 Proc. Nat. Acad. Sci. 104 36
[23] Zachary WW, 1977 J. Anthropol. Res. 33
[24] (Cornell KDD Cup)
[25] Hoerdt M and Magoni D, 2003 Proc. 11th Int. Conf. on Software, Telecommunications and Computer
Networks p 257
[26] Albert R, Jeong H and Barab´asi A-L, 1999 Nature 401 130
[27] (Laboratory for Web Algorithmics)
[28] (Stanford WebBase Project)
[29] Danon L, D´ıaz-Guilera A, Duch J and Arenas A, 2005 J. Stat. Mech. P09008
[30] Reichardt J and Bornholdt S, 2004 Phys.Rev.Lett.93 218701
[31] Duch J and Arenas A, 2005 Phys. Rev. E 72 027104
[32] Lancichinetti A, Fortunato S and Kertesz J, 2008 arXiv:0802.1218
[33] Lambiotte R, Blondel V D, de Kerchove C, Huens E, Prieur C, Smoreda Z and Van Dooren P, 2008
Physica A 387 5317
[34] Palla G, Barab´asi A-L and Vicsek T, 2007 Nature 446 664
[35] Onnela J-P, Saram¨aki J, Hyv¨onen J, Szab´oG,LazerD,KaskiK,Kert´esz J and Barab´asi A-L, 2007 Proc.
Nat. Acad. Sci. 104 7332
doi:10.1088/1742-5468/2008/10/P10008 11
J. Stat. Mech.
(2008) P10008
Fast unfolding of communities in large networks
[36] Lambiotte R, Ausloos M and Holyst J A, 2007 Phys. Rev. E 75 030101(R)
[37] Arenas A, Fern´andez A and G´omez S, 2008 New J. Phys. 10 053039
[38] Delvenne J-C, Yaliraki S and Barahona M, 2008 in preparation
doi:10.1088/1742-5468/2008/10/P10008 12
... where ω ij is the weight of the edge going from node i to node j, m is the sum over all weights in the network, k x is the sum of all weights of the edges at node i, and c i is the community node i is assigned to. In the case of a weighted graph, the goal of community detection can be reformulated to the problem of maximizing the modularity of a given graph [28]. ...
... This is an agglomerative clustering method, as communities are formed by merging similar nodes, and thus it is an unsupervised data mining method. The authors of the algorithm suggest that the complexity is linear on sparse graphs [28]. ...
... Otherwise, the node i stays in its community. This is applied repeatably for all nodes until no further gain in modularity can be achieved [28]. In the second phase, a new graph is created, having the communities formed in phase one as its nodes. ...
Full-text available
A deep understanding about a field of research is valuable for academic researchers. In addition to technical knowledge, this includes knowledge about subareas, open research questions, and social communities (networks) of individuals and organizations within a given field. With bibliometric analyses, researchers can acquire quantitatively valuable knowledge about a research area by using bibliographic information on academic publications provided by bibliographic data providers. Bibliometric analyses include the calculation of bibliometric networks to describe affiliations or similarities of bibliometric entities (e.g., authors) and group them into clusters representing subareas or communities. Calculating and visualizing bibliometric networks is a nontrivial and time-consuming data science task that requires highly skilled individuals. In addition to domain knowledge, researchers must often provide statistical knowledge and programming skills or use software tools having limited functionality and usability. In this paper, we present the ambalytics bibliometric platform, which reduces the complexity of bibliometric network analysis and the visualization of results. It accompanies users through the process of bibliometric analysis and eliminates the need for individuals to have programming skills and statistical knowledge, while preserving advanced functionality, such as algorithm parameterization, for experts. As a proof-of-concept, and as an example of bibliometric analyses outcomes, the calculation of research fronts networks based on a hybrid similarity approach is shown. Being designed to scale, ambalytics makes use of distributed systems concepts and technologies. It is based on the microservice architecture concept and uses the Kubernetes framework for orchestration. This paper presents the initial building block of a comprehensive bibliometric analysis platform called ambalytics, which aims at a high usability for users as well as scalability.
... These groups are deemed "modules" or "communities" [93]. Among the diverse algorithms of community detection, this study used the algorithm proposed by Blondel et al. [94]. This algorithm follows a local optimization method suggested by Girvan and Newman [93] to investigate some partitions. ...
... This step is iterated until the value of modularity no longer increases. The anymore modularity ranges from −1 to 1 [94]. The algorithm is available in the network analysis and visualization program for Gephi; thus, the community detection was performed by Gephi version 0.9.7. ...
Full-text available
Developing bidirectional urban networks within areas in megacities is an essential spatial strategy across regions today. In 2018, Korea began its Bu-Ul-Gyeong (BUG) megacity project. Today, Korea is working to improve functional polycentric urban networks within the BUG megacity. To uncover insights useful for this project, this study sought to examine urban network patterns (e.g., network asymmetries and imbalances in the sizes and directions of their weighted flows) and identify the primary and secondary centers of the BUG megacity using mobile flow data from 2019 to 2020. Specifically, a three-step social network analysis was conducted across different geographical scales; namely: (1) the BUG megacity, (2) South Gyeongsang Province (SGP), and (3) every community in SGP. Eigenvector centrality and flow betweenness centrality revealed two primary centers (Changwon and Jinju) and four secondary centers (Haman, Sacheon, Tongyeong, and Geochang). Unidirectional and hierarchical connections were evident between the primary and secondary centers. In response to these findings, this paper proposes some beneficial strategies for the region’s public transportation networks to prevent small- and medium-sized cities from being marginalized and to enhance horizontal urban connectivity in megacities.
... The DeltaCon similarity score 16 exhibits the properties of edge importance (changes leading to disconnected graph penalized more than the ones that maintain connectivity), weight awareness (the larger the weight of the removed edge, the greater the impact on similarity score), edge-"submodularity" (a speci c change in sparse network is more important than in a much denser equally sized network) and focus awareness (random changes are less important than targeted changes of the same extent). The similarity score varies between 0 and 1, where 0 means totally different graphs and 1 means identical graphs, and is more robust than other network comparison metrics 29 . ...
Full-text available
Structural covariance network (SCN) studies on first-episode antipsychotic-naïve psychosis (FEAP) have examined less granular parcellations on one morphometric feature reporting lower network resilience among other findings. We examined SCNs of volumes, cortical thickness, and surface area using the Human Connectome Project atlas-based parcellation of 358 regions from 79 FEAP and 68 controls to comprehensively characterize the networks using descriptive and perturbational network neuroscience approach. Using graph theoretic methods, we examined network integration, segregation, centrality, community structure, and hub distribution across small-worldness threshold range and correlated them with psychopathology severity. We used simulated nodal “attacks” (removal of nodes and all their edges) to investigate network resilience, and calculated DeltaCon similarity scores and contrasted the removed nodes to characterize the impact of simulated attacks. Compared to controls, FEAP SCN showed higher betweenness centrality (BC) and lower degree in all three morphometric features and disintegrated with fewer attacks with no change in global efficiency. SCNs showed higher similarity score at the first point of disintegration with ≈54% top-ranked BC nodes attacked. FEAP communities consisted of fewer prefrontal, auditory and visual regions. Lower BC, and higher clustering and degree were associated with greater positive and negative symptom severity. Negative symptoms required twice the changes in these metrics. Globally sparse but locally dense network with more higher-importance nodes in FEAP could result in higher communication cost compared to controls. FEAP network disintegration with fewer attacks suggests lower resilience without altering efficiency measure. Greater network disarray underlying negative symptom severity possibly explains the therapeutic challenge.
... Aun con todo, en los sistémicos sociales, la modularidad no siempre es clara, pero eso no impide que se estimen o sea lógico deducir los componentes de un sistema social, pero dependen del observador y de la modelación que hace del fenómeno observado. Blondel et al. (2008), y el crédito del cálculo de resolución, que acompaña al anterior, es de Lambiotte, Delvenne y Barahona (2008). En resumidas cuentas, el algoritmo y la resolución están pensados para estimar la fuerza de división de una red en módulos. ...
Full-text available
Las instituciones son fundamentales para delimitar la sociedad, ya que permiten la legitimación de valores y principios desde las afinidades que puedan tener los individuos que la componen. Teniendo en cuenta lo anterior, en este trabajo se explora la hipótesis de que estudiantes universitarios con valores similares son afines entre sí. Para probarla se aplican dos instrumentos: primero, el Estudio de valores de Allport, Vernon y Lindzey, y segundo, una encuesta de percepción de afinidad de elaboración propia. Esto con la finalidad de explorar la formación de microsociedades, así como del rol que cumplen los valores e instituciones mediante un análisis de afinidad de valores en estudiantes universitarios. A partir de un procesamiento de modularidad se examinan las relaciones sociales de dos grupos de estudiantes que cursan una carrera universitaria y se compara su composición de valores. La evidencia permite concluir que no existe una composición de valores sustancialmente distinta en estudiantes, por lo que sus valores pueden surgir de un contexto más allá del universitario, por ejemplo, su localidad, región o un fenómeno generacional. Los resultados señalan que las afinidades de los participantes fueron recíprocas, indicando que las percepciones entre ellos coinciden, por lo que se abona a la noción de conciencias supraindividuales.
... The Community analysis (component analysis) then identifies the number of different components (in the case of community modularity) in the network based on the modularity detection analysis (Blondel, 2008), as follows: ...
Conference Paper
Full-text available
The 21st century has seen environmental degradation as one of the main challenges experienced today. Syria, which is in a peculiar situation due to the ongoing war, is not exempt from this predicament. The objective of this article is to examine the relationship among carbon dioxide, electricity consumption, and economic growth for the time 1971 to 2014. In reaching this objective, the paper applied a cointegration test (which was done after the ARDL long- run bounds Test) and found the existence of cointegration amongst the variables, which was significant and adjusted to long-run equilibrium at -0.94 as expected consumption of electricity and economic growth all have a positive increment towards carbon dioxide emissions. These raise concern amongst policymakers who need to build environmentally friendly means of energy use and economic activities.
A fundamental fact about mutualisms is that these mutually beneficial interactions often harbor cheaters that benefit from the use of resources and services without providing any positive feedback to other species. The role of cheaters in the evolutionary dynamics of mutualisms has long been recognized, yet their broader impacts at the community level, beyond species they directly interact with, is still poorly understood. Because mutualisms form networks often involving dozens of species, indirect effects generated by cheaters may cascade through the whole community, reshaping trait evolution. Here, we study how cheating interactions can influence coevolution in mutualistic networks. We combined a coevolutionary model, empirical data on animal–plant mutualistic networks and numerical simulations to show that high trait disparity emerges as a consequence of the negative effect of cheaters on victim fitness, which in turn fuels selection favoring victim traits that are increasingly different from the cheaters' traits. Intermediate levels of cheating interactions in a network can lead to the formation of groups of species phenotypically similar to each other and distinct from species in other groups, generating clustered trait patterns. The resulting clustered trait pattern, in turn, changes the pattern of interaction in simulated networks, fostering the formation of modules of interacting species and reducing nestedness. Our results indicate that directional selection imposed by cheaters on their victims counteracts selection for trait convergence imposed by mutualists, leading to the emergence of modules of phenotypically similar interacting species but phenotypically distinct from other modules. Based on these results, we suggest that cheaters might be a fundamental missing element for our understanding of how multispecies selection shapes the trait distribution and structure of mutualistic networks.
Full-text available
This paper examines the communicational and linguistic mechanisms associated with the circulation of conspiracy-oriented discourse on Twitter, based on a corpus of more than 55.5 million tweets in French about Covid-19 vaccination. Using an innovative tensor decomposition method, the paper identifies clusters of accounts structured over time around specific events relayed in the media. Quantitative and qualitative analyses of tweets from a specific cluster highlight the influence on its circulation of the constituents of a particular tweet, and of the constructional and pragma-semantic level of the hashtags which are associated with the cluster
Full-text available
Determining the repertoire of a microbe's molecular functions is a central question in microbial genomics. Modern techniques achieve this goal by comparing microbial genetic material against reference databases of functionally annotated genes/proteins or known taxonomic markers such as 16S rRNA. Here we describe a novel approach to exploring bacterial functional repertoires without reference databases. Our Fusion scheme establishes functional relationships between bacteria and thus assigns organisms to Fusion taxa that differ from otherwise defined taxonomic clades. Three key findings of our work stand out. First, Fusion profile comparisons outperform existing functional annotation schemes in recovering taxonomic labels. Second, Fusion-derived functional co-occurrence profiles reflect known metabolic pathways, suggesting a route for discovery of new ones. Finally, our alignment-free nucleic acid-based Siamese Neural Network model, trained using Fusion functions, enables finding shared functionality of very distant, possibly structurally different, microbial homologs. Our work can thus help annotate functional repertoires of bacterial organisms and further guide our understanding of microbial communities.
Discovering communities in networks constitutes an active and significant research field, especially in real-world networks, where communities usually represent functional units of these networks. In the introduced work, we study the community discovery problem in real-world networks from a different perspective. Our motivation stems from inquiries related to the cause of a node’s participation in a community. Therefore, we incorporate the concepts of causality and responsibility into a novel framework. Our goal is to set up a framework to discover explainable causal relations between nodes and communities, which is a topic that has not been investigated before. We evaluate the proposed framework by experimenting on real-world networks. Moreover, the introduced framework is flexible to be used to other network-related analysis tasks, beyond community detection, such as link prediction, anomaly detection, influence maximization.
We show that inner composition alignment networks derived for financial time-series data, studied in response to worldwide lockdown imposed in response to COVID-19 situation, show distinct patterns before, during and after lockdown phase. It is observed that significant couplings between companies as captured by inner composition alignment between time series, reduced considerably across the globe during lockdown and post-lockdown recovery period. The study of global community structure of the networks show that factions of companies emerge during recovery phase, with strong coupling within the members of the faction group, a trend which was absent before lockdown period. The study of strongly connected components of the networks further show that market has fragmented in response to COVID-19 situation. We find that most central firms as characterized by in-degree, out-degree and betweenness centralities belong to Chinese and Japanese economies, indicating a role played by these organizations in financial information propagation across the globe. We further observe that recovery phase of the lockdown period is strongly influenced by financial sector, which is one of the main result of this study. It is also observed that two different group of companies, which may not be co-moving, emerge across economies during COVID-19. We further notice that many companies in US and European economy tend to shield themselves from local influences.
Full-text available
Despite its increasing role in communication, the world wide web remains the least controlled medium: any individual or institution can create websites with unrestricted number of documents and links. While great efforts are made to map and characterize the Internet's infrastructure, little is known about the topology of the web. Here we take a first step to fill this gap: we use local connectivity measurements to construct a topological model of the world wide web, allowing us to explore and characterize its large scale properties. Comment: 5 pages, 1 figure, updated with most recent results on the size of the www
Full-text available
Data from a voluntary association are used to construct a new formal model for a traditional anthropological problem, fission in small groups. The process leading to fission is viewed as an unequal flow of sentiments and information across the ties in a social network. This flow is unequal because it is uniquely constrained by the contextual range and sensitivity of each relationship in the network. The subsequent differential sharing of sentiments leads to the formation of subgroups with more internal stability than the group as a whole, and results in fission. The Ford-Fulkerson labeling algorithm allows an accurate prediction of membership in the subgroups and of the locus of the fission to be made from measurements of the potential for information flow across each edge in the network. Methods for measurement of potential information flow are discussed, and it is shown that all appropriate techniques will generate the same predictions.
Full-text available
Modular structure is ubiquitous in real-world complex networks, and its detection is important because it gives insights into the structure–functionality relationship. The standard approach is based on the optimization of a quality function, modularity, which is a relative quality measure for the partition of a network into modules. Recently, some authors (Fortunato and Barthélemy 2007 Proc. Natl Acad. Sci. USA 104 36 and Kumpula et al 2007 Eur. Phys. J. B 56 41) have pointed out that the optimization of modularity has a fundamental drawback: the existence of a resolution limit beyond which no modular structure can be detected even though these modules might have their own entity. The reason is that several topological descriptions of the network coexist at different scales, which is, in general, a fingerprint of complex systems. Here, we propose a method that allows for multiple resolution screening of the modular structure. The method has been validated using synthetic networks, discovering the predefined structures at all scales. Its application to two real social networks allows us to find the exact splits reported in the literature, as well as the substructure beyond the actual split.
Full-text available
We present a method that allows for the discovery of communities within graphs of arbitrary size in times that scale linearly with their size. This method avoids edge cutting and is based on notions of voltage drops across networks that are both intuitive and easy to solve regardless of the complexity of the graph involved. We additionally show how this algorithm allows for the swift discovery of the community surrounding a given node without having to extract all the communities out of a graph.
Conference Paper
Full-text available
Community analysis algorithm proposed by Clauset, New- man, and Moore (CNM algorithm) finds community struc- ture in social networks. Unfortunately, CNM algorithm does not scale well and its use is practically limited to networks whose sizes are up to 500,000 nodes. We show that this inef- ficiency is caused from merging communities in unbalanced manner and that a simple heuristics that attempts to merge community structures in a balanced manner can dramati- cally improve community structure analysis. The proposed techniques are tested using data sets obtained from exist- ing social networking service that hosts 5.5 million users. We have tested three three variations of the heuristics. The fastest method processes a SNS friendship network with 1 millionusersin5 minutes(70 timesfasterthan CNM)and an- other friendship network with 4 million users in 35 minutes, respectively. Another one processes a network with 500,000 nodes in 50 minutes (7 times faster than CNM), finds com- munity structures that has improved modularity, and scales to a network with 5.5 million. Further detail is reported in (3).
In this paper, we analyze statistical properties of a communication network constructed from the records of a mobile phone company. The network consists of 2.5 million customers that have placed 810 million communications (phone calls and text messages) over a period of 6 months and for whom we have geographical home localization information. It is shown that the degree distribution in this network has a power-law degree distribution k−5k−5 and that the probability that two customers are connected by a link follows a gravity model, i.e. decreases as d−2d−2, where dd is the distance between the customers. We also consider the geographical extension of communication triangles and we show that communication triangles are not only composed of geographically adjacent nodes but that they may extend over large distances. This last property is not captured by the existing models of geographical networks and in a last section we propose a new model that reproduces the observed property. Our model, which is based on the migration and on the local adaptation of agents, is then studied analytically and the resulting predictions are confirmed by computer simulations.
Many complex systems in nature and society can be described in terms of networks capturing the intricate web of connections among the units they are made of1, 2, 3, 4. A key question is how to interpret the global organization of such networks as the coexistence of their structural subunits (communities) associated with more highly interconnected parts. Identifying these a priori unknown building blocks (such as functionally related proteins5, 6, industrial sectors7 and groups of people8, 9) is crucial to the understanding of the structural and functional properties of networks. The existing deterministic methods used for large networks find separated communities, whereas most of the actual networks are made of highly overlapping cohesive groups of nodes. Here we introduce an approach to analysing the main statistical features of the interwoven sets of overlapping communities that makes a step towards uncovering the modular structure of complex systems. After defining a set of new characteristic quantities for the statistics of communities, we apply an efficient technique for exploring overlapping communities on a large scale. We find that overlaps are significant, and the distributions we introduce reveal universal features of networks. Our studies of collaboration, word-association and protein interaction graphs show that the web of communities has non-trivial correlations and specific scaling properties.
Conference Paper
Dense subgraphs of sparse graphs (communities), which appear in most real-world complex networks, play an important role in many con- texts. Computing them however is generally expensive. We propose here a measure of similarity between vertices based on random walks which has several important advantages: it captures well the community structure in a network, it can be computed efficiently, and it can be used in an ag- glomerative algorithm to compute efficiently the community structureof a network. We propose such an algorithm, called Walktrap, which runs in time O(mn2) and space O(n2) in the worst case, and in time O(n2 log n) and space O(n2) in most real-world cases (n and m are respectively the number of vertices and edges in the input graph). Extensive comparison tests show that our algorithm surpasses previously proposed ones concern- ing the quality of the obtained community structures and that it stands among the best ones concerning the running time.
Dense subgraphs of sparse graphs (communities), which appear in most real-world complex networks, play an important role in many contexts. Computing them however is generally expensive. We propose here a measure of similarities between vertices based on random walks which has several important advantages: it captures well the community structure in a network, it can be computed efficiently, and it can be used in an agglomerative algorithm to compute efficiently the community structure of a network. We propose such an algorithm, called Walktrap, which runs in time O(mn^2) and space O(n^2) in the worst case, and in time O(n^2log n) and space O(n^2) in most real-world cases (n and m are respectively the number of vertices and edges in the input graph). Extensive comparison tests show that our algorithm surpasses previously proposed ones concerning the quality of the obtained community structures and that it stands among the best ones concerning the running time.