ArticlePDF Available

ANGEL: efficient, and effective, node-centric community discovery in static and dynamic networks


Abstract and Figures

Abstract Community discovery is one of the most challenging tasks in social network analysis. During the last decades, several algorithms have been proposed with the aim of identifying communities in complex networks, each one searching for mesoscale topologies having different and peculiar characteristics. Among such vast literature, an interesting family of Community Discovery algorithms, designed for the analysis of social network data, is represented by overlapping, node-centric approaches. In this work, following such line of research, we propose Angel, an algorithm that aims to lower the computational complexity of previous solutions while ensuring the identification of high-quality overlapping partitions. We compare Angel, both on synthetic and real-world datasets, against state of the art community discovery algorithms designed for the same community definition. Our experiments underline the effectiveness and efficiency of the proposed methodology, confirmed by its ability to constantly outperform the identified competitors.
Content may be subject to copyright.
Applied Network Science
Rossetti Applied Network Science (2020) 5:26
RESEARCH Open Access
ANGEL: efficient, and effective,
node-centric community discovery in static
and dynamic networks
Giulio Rossetti
KDD-Lab, ISTI-CNR, Pisa, Italy
Community discovery is one of the most challenging tasks in social network analysis.
During the last decades, several algorithms have been proposed with the aim of
identifying communities in complex networks, each one searching for mesoscale
topologies having different and peculiar characteristics. Among such vast literature, an
interesting family of Community Discovery algorithms, designed for the analysis of
social network data, is represented by overlapping, node-centric approaches. In this
work, following such line of research, we propose ANGEL, an algorithm that aims to
lower the computational complexity of previous solutions while ensuring the
identification of high-quality overlapping partitions. We compare ANGEL,bothon
synthetic and real-world datasets, against state of the art community discovery
algorithms designed for the same community definition. Our experiments underline
the effectiveness and efficiency of the proposed methodology, confirmed by its ability
to constantly outperform the identified competitors.
Keywords: Complex networks, Dynamic networks, Community discovery
Community discovery (henceforth CD), the task of decomposing a complex network
topology into meaningful node clusters, is one of the oldest and most discussed problems
in complex network analysis (Coscia et al. 2011;Fortunato2010). One of the main reasons
behind the attention it has received during the last decades lies in its intrinsic complex-
ity, strongly tied to its overall ill-posedness. Indeed, complex networks researchers agree
that it is not possible to provide a single and unique formalization that covers all the pos-
sible characteristics a community partition may satisfy. Usually, every CD approach is
designed to provide a different point of view on how to partition a graph: in this scenario,
the solutions proposed by different authors were often proven to perform well when spe-
cific assumptions can be made on the analyzed topology. Nonetheless, decomposing a
complex structure in a set of meaningful components represents per se a step required
by several analytical tasks. Such peculiarity has lead to the definition of several “meta”
community definitions, often tied to specific analytical needs. For instance, classic works
© The Author(s). 2020 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate
credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were
made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless
indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your
intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this licence, visit
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 2 of 23
intuitively describe communities as sets of nodes closer among them than with the rest of
the network, while others, looking at the same problem from another angle, only define
such topologies as dense network subgraphs. A general, high-level, formulation of the
Community Discovery problem definition is the following:
Definition 1 (Community Discovery (CD)) Given a network G, a community C is
defined as a set of nodes in G: C ={v1,v2,...,vn}. The community discovery problem aims
to identify the set Cof all the communities in G.
The absence of a unique, well-posed, definition of what a community in a complex
network should represent is only one of the issues to face when approaching network
clustering. Indeed, the evolution through time of a network topology plays a major role
in the way communities can be defined and extracted. Even though the CD problem
has been classically studied considering the underlying network topology as “frozen in
time", recently a novel branch of research addressed the problem of studying the depen-
dant evolution of networks and their communities. Complex networks are often used to
model dynamic objects – e.g., social phenomena, economic transactions, human mobil-
ity – composed by nodes and edges that may appear and vanish as time goes by. When
considering this temporally enriched scenario, we need to revise the formulation of the
classical Community Discovery problem. We will then talk of Dynamic Community Dis-
covery (henceforth referred as DCD), a problem that can be defined by abstracting the
specific CD definition as done in Rossetti and Cazabet (2018):
Definition 2 (Dynamic Community Discovery (DCD)) Given a dynamic network DG,
a Dynamic Community DC is defined as a set of (node, periods) pairs:
DC ={(v1,P1),(v2,P2),...,(vn,Pn)},withP
n=((ts0,te0),(ts1,te1)...(tsN ,teN )),with
tste. Dynamic Community Discovery aims to identify the set Cof all the dynamic
communities in DG.
Both proposed problem meta-definitions allow multiple solutions to the network
clustering under different constraints. As an example, such definitions do not explic-
itly require complete coverage of the nodes, nor specify if the identified clustering
represents a neat nodes partition or, instead, a cover (thus allowing overlaps among
communities). In this work, we introduce a CD algorithm, ANGEL, tailored to extract
overlapping communities from a complex network. Our approach is primarily designed
for social networks analysis and belongs to a well-known subfamily of Community
Discovery approaches often identified by the keywords bottom-up and node-centric
(Rossetti et al. 2017b). ANGEL aims to provide a fast way to compute reliable over-
lapping network partitions in the absence of topology dynamics. However, as we
underlined, the unfolding of time plays a significant role in the structures describ-
ing social phenomena. To cope with such intrinsic evolution, we leverage ANGEL to
design a simple Dynamic Community Discovery approach that can be used to track
dynamic communities and their life-cycles. Both the proposed approaches focus on low-
ering the computational complexity of existing methods proposing scalable sequential –
although, easily parallelizable – solutions to a very demanding task: overlapping network
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 3 of 23
The paper is organized as follows. “Related works” section covers the relevant liter-
ature on community discovery needed to frame the proposed approach. In “ANGEL:
static community discovery” section we introduce our static node-centric algorithm,
ANGEL. There we discuss its rationale, the properties it holds as well as its computa-
tional complexity. In “ANGEL evaluation” section we evaluate the proposed method on
both synthetic and real-world datasets for which ground truth communities are known
in advance. To better discuss the resemblance of ANGEL partitions to ground truth
ones as well as its execution times, we compare the proposed method with state-of-
art competitors sharing the same rationale. In “ANGEL on dynamic networks”section
we introduce the extension of ANGEL we designed to cope with dynamic network
topologies. There we frame the proposed method in its general class and discuss its
computational complexity. In “Dynamic ANGEL evaluation” section, as done for ANGEL,
we evaluate its extension on both synthetic benchmarks and real-world dynamic net-
works. To do so, the concept of community life-cycle is introduced, and qualitative
analysis of community event trends is performed. Finally, “Conclusion” section concludes
the paper.
Related works
Community discovery is a widely discussed and studied problem. Researchers contin-
uously propose novel approaches with the aim of solving specific declinations of this
complex, and ill-posed, problem. Due to the massive literature available in this field,
several attempts were made to organize and cluster methods identifying some common
grounds. Among the others, the surveys of Fortunato (2010) and Fortunato and Hric
(2016)andCoscia(Cosciaetal.2011) propose complete, detailed and extensive tax-
onomies for classic algorithms. However, due to the intrinsic complexity of the problem,
several thematic surveys emerged, each focusing on a different declination (for instance
considerint overlapping (Xie et al. 2013), directed (Malliaros and Vazirgiannis 2013),
node-centric (Rossetti et al. 2017b)aswellasdynamiccommunitydiscovery(Cazabet
et al. 2017;RossettiandCazabet2018)).
Static Community Discovery. The algorithmic solutions we propose share a very spe-
cific goal: identify overlapping network partitions following a bottom-up, node-centric,
strategy. Such an approach is often adopted while analyzing social network contexts
(Rossetti et al. 2015;2016; Milli et al. 2015), scenarios in which it is important to
take into account the individual perspective on their local communities. Most impor-
tantly, in social scenarios, neat partitions are rarely semantically coherent or easily
identifiable. Following such a rationale, ANGEL leverages individual ego-networks to
access the node-centric perspective of the analyzed social graph. The growing availabil-
ity of social media data has indeed allowed for extensive studies of such ego-centered
topologies: among them in (Arnaboldi et al. 2017) Facebook and Twitter datasets were
studied to relate online and offline properties of ego-networks. Such procedure, origi-
nally proposed in Coscia et al. (2012)andCosciaetal.(2014a) were also extended to
parallel implementations, as in Amoretti et al. (2016), and generalized in a high-level
framework (Soundarajan and Hopcroft 2015). Moreover, several approaches leverage the
concept of ego-network to design heterogeneous community definitions (Epasto et al.
2017;Buzunetal.2014). Other common strategies to design node-centric approaches,
avoiding the use of ego-networks, are the seed set expansion (Moradi et al. 2014;
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 4 of 23
Whang et al. 2016), and community diffusion ones (Kumpula et al. 2008;Raghavan
et al. 2007). The former decompose the community detection into two steps: iden-
tification of the seed nodes and definition of an iterative rule that describe how
they attract nodes to form communities around them. Conversely, the latter let each
node in the graph to autonomously chose its community by observing the choices
made by its neighborhood. A classic example of this family of approaches is offered
by the Label Propagation algorithm used by ANGEL to identify local communities
(Raghavan et al. 2007).
Dynamic Community Discovery. Indeed, a significant number of systems can
be modeled as temporal networks: cellular processes, social communications, large
infrastructures (i.e., call graphs and web graphs) posses both network and tempo-
ral aspects that make them a perfect fit for dynamic network modeling. One of the
first works underlining the needs for a dedicated framework for analyzing evolv-
ing network structure is indeed (Holme and Saramäki 2012). Several formalisms
for representing evolving networks have been proposed to support the definition of
such revised analytical framework: Temporal Networks (Holme and Saramäki 2012),
Time-Varying Graphs (Casteigts et al. 2012), Interaction Networks (Rossetti et al.
2016), and Link Streams (Viard et al. 2016), to name the most famous. Leverag-
ing such temporally enriched models novel community discovery approaches started
taking into account the temporal dimension, following different strategies. In Ros-
setti and Cazabet (2018), three families of DCD algorithms are identified and
Instant-optimal CD
assumes that communities existing at
only depend on the
state of the network at
, as done by our dynamic ANGEL extension.
Temporal Trade-off CD
assumes that communities defined at an instant
do not
only depend on the topology of the network at that time, but also on the
evolutions of the topology,
partitions found, or both.
- Finally,
Cross-Time CD
shifts the from searching communities relevant at a
particular time to searching communities relevant when considering the whole
network evolution.
Within the first family are grouped several two-steps, Identify&Match, algorithms. The
common ground of such approaches, e.g., (Palla et al. 2007; Takaffoli et al. 2011;Morini
et al. 2017), is that they are easily parallelizable while suffering from some instability
due to the matching phase performed as post-processing. Conversely, Temporal Trade-
off approaches focus on smoothly identifying community evolutions as they happen.
Such algorithms, e.g., (Cazabet et al. 2010;ZakrzewskaandBader2015;Rossettietal.
2017a) , are designed to deal with high-frequency node interactions, are not easily par-
allelizable and prone to “avalanche effects" (i.e., since they focus on local community
perturbations the node groups tend, as time goes by, to increase their sizes). Finally,
algorithms of the latter family search for stable communities across time, e.g., (Matias
and Miele 2016;Muchaetal.2010; Himmel et al. 2016). They often work upon tempo-
ral network aggregations built leveraging the complete knowledge of nodes and edges
evolution. As a result, they are usually neither easily parallelizable nor applicable in
online scenarios.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 5 of 23
Input:G:(V,E), the graph; φ, the merging threshold.
Output:Ca set of overlapping communities.
1for vVdo // Step #1
2eEgoMinusEgo(v, G); // Step #2
3C(v)LabelPropagation(e) ; // Step #3
5ncoms = |C|
6acoms = 0
7while ncoms != acoms do // Step #4
8acoms = ncoms
9CDecreasingSizeSorting(C);// Step #5
10 for cCdo
11 CPrecisionMerge(c, C,φ); // Step #6
12 ncoms = |C|
13 return C
ANGEL: static community discovery
In this section, we present our bottom-up solution to the community discovery problem.
In “Algorithm rationale” section we discuss the core of ANGEL1. Our approach follows a
well-known pattern composed by two phases: i) construction of local communities mov-
ing from ego-network structures and, ii) definition of mesoscale topologies by aggregating
the identified local-scale ones. Moreover, in “Properties”and“Complexity”sectionswe
discuss the properties of the proposed algorithm and provide bounds to its complexity.
Algorithm rationale
The algorithmic schema of ANGEL is borrowed from the one firstly adopted in
(Coscia et al. 2012) where the authors propose DEMON an approach whose main goal
was to identify local communities by capturing individual nodes perspectives on their
neighbourhoods and using them to build mesoscale ones. ANGEL follows the same
rationale: however, conversely from its predecessor, it focuses on lowering the time com-
plexity while at the same time increasing the partition quality (as will be discussed
in “Complexity”section).
ANGEL starts taking as input a graph G, a merging threshold φand an empty set of com-
munities C. The algorithm main loop cycles over each node, so to generate all the possible
points of view of the network structure and guarantee complete coverage of its overall
topology (Step #1 in Algorithm 1). To do so, for each node v, our algorithm applies the
EgoMinusEgo(v,G)(Step #2 in Algorithm 1) operation as defined in Coscia et al. (2014b).
Such function extracts the ego-network centred in the node v– e.g., the graph induced
on Gand built upon vand its first order neighbours – then removes vfrom it, obtaining
a novel, filtered, graph substructure. ANGEL removes vsince, by definition, it is directly
linked to all nodes in its ego-network, connections that would lead to noise in the iden-
tification of local communities. A single node connecting the entire sub-graph will make
1Code available at: and within the CDlib python library (Rossetti et al. 2019)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 6 of 23
ALGORITHM 2: PrecisionMerge
Input:x,acommunity;C, a set of overlapping communities; φ, the merging threshold.
Output:C, a set of overlapping communities.
1com_to_freq community_frequency(x) ; // Step #A
2for com, freq com_to_freq do
3if freq
|x|φthen // Step #B
6return C
all nodes very close, even if they are not in the same local community. Once obtained the
ego-minus-ego graph, the next step is to compute the local communities it contains (Step
#3 in Algorithm 1). The algorithm performs this step by using a CD approach borrowed
from the literature: Label Propagation (LP)(Raghavan et al. 2007). This choice, already
adopted in (Coscia et al. 2012), has been made for the following reasons:
1. LP is known as the least complex algorithm in the literature, reaching a quasi-linear
time complexity in terms of nodes. However,
2. LP will return results of a quality comparable to more complex algorithms(Coscia
et al. 2011).
Reason #1 is particularly important since Step #3 of our pseudocode needs to be per-
formed once for every node of the network, thus making it unacceptable to spend a
super-linear time for each node. Notice that instead of LP any other community discovery
algorithm (both overlapping or not) can be used (impacting both on the algorithmic com-
plexity and partition quality). Given the linear complexity (in the number of nodes of the
extracted ego-minus-ego graph) of Step #3, we refer to this as the inner loop for finding
the local communities. Due to the importance of LP for our approach and to shed lights on
how it works, we briefly describe its classical formulation (Raghavan et al. 2007). Suppose
that a node vhas neighbors v1,v2, ..., vkand that each one of them carries a label denot-
ing the community it belongs. Then, during each iteration, the label of vis updated to
the majority label of its neighbours. As the labels propagate, densely connected groups of
nodes quickly reach a consensus on a unique label. At the end of the propagation process,
nodes sharing the same labels identify the resulting communities.
In case of bow-tie situations – e.g., a node having an equal maximum number of neigh-
bors in two or more communities (example in Fig. 1) – the classic definition of the LP
algorithm randomly select a single label for the contended node. ANGEL,conversely,han-
dle this situation – that otherwise can led to nondeterministic behaviours – by allowing
soft community memberships: each node can thus belong to multiple communities in
case of bow-tie configuration. The result of Steps #1-3 of Algorithm 1 is a set of local
communities C(v), according to the perspective of a specific node, v, of the network. Dif-
ferently, from what done in DEMON,ANGEL does not reintroduce the ego in each local
community to reduce the noisy effects hubs play during the merging step. Since local
communities can be seen as a biased and partial view of the real community structure of
G, the result of ANGEL needs further processing: namely, a merging step that simplifies
the local partition present in C.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 7 of 23
Fig. 1 Label Propagation. In the example, the node zrepresents the center of a bow-tie structure since it has
its neighbors divide equally among two different communities ({u,v} belonging the red one, {k,w}tothe
yellow one). In such scenario the classic LP formulation will assign zto only one of the two communities after
experiencing a ping-pong effect: this will cause nondeterminism. Our approach instead will assign zto both
partitions producing a soft, but stable, community membership
Once the outer loop on the network nodes is completed, ANGEL leverages the PRECI-
SIONMERGE function to compact the community set Cso to avoid the presence of fully
contained communities in it. Such function (Step #6, detailed in Algorithm 2) implements
a deterministic merging strategy and is applied iteratively until reaching convergence
(Step #4) – e.g., until the communities in Ccannot be merged further. To assure that all the
possible community merges are performed at each iteration Cis ordered from the smallest
community to the biggest (Algorithm 1, #Step 6). This merging step is a crucial one since
it needs to be repeated for each one of the local communities. In DEMON such operation
requires the computation for each pair of communities (x,y),xC(v)and yC,ofan
overlap measure (i.e. Jaccard index) and to evaluate if it overcomes a user defined thresh-
old. This approach, although valid, has a major drawback: given a community xC(v)it
requires O(|C|)evaluations to identify its best match among its peers. Indeed, such kind
of strategy represents a costly bottleneck requiring an overall O(|C|2)complexity while
applied to all the identified local communities. ANGEL aims to drastically reduce such
computational complexity by performing the matches leveraging a greedy strategy. To do
so, it proceeds in the following way:
i) ANGEL assumes that each node carries, as additional information, the identifiers of
all the communities in Cit already belongs to;
ii) in Step #A (Algorithm 2) for each local community
is computed the frequency of
the community identifiers associated with its nodes;
iii) in Step #B, for each pair (community_id,
) is computed its Precision w.r.t.
, namely the percentage of nodes in
that also belong to community_id;
iv) iff the precision ratio is greater (or equal) than a given threshold φthe local
is merged with community_id: their union is added to Cand the
original communities are removed from the same set.
Operating in this way it is avoided the time expensive computation of community inter-
sections required by Jaccard-like measures since all the containment testing can be done
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 8 of 23
in place. Figure 2shows two examples of Angel clustering of the Zachary Karate club
network obtained varying the φthreshold. As expected, increasing the ψthreshold, we
obtain a higher number of communities since lower quality merges cannot take place.
The proposed approach posses two nice properties: it produces a deterministic out-
put (given the as input a network Gand a threshold φ), and it allows for a parallel
Property 1 (Determinism)There exists a unique C=ANGEL(G,φ) for any given G and
φ, disregarding the order of visit of the nodes in G.
To prove the determinism of ANGEL it is mandatory to break its execution in two well-
defined steps: (i) local community extraction and (ii) merging of local communities.
i) Local communities: Label Propagation identifies communities by applying a greedy
strategy. In its classical formulation (Raghavan et al. 2007) it does not assure
convergence to a stable partition due to the so-called “label
problem" (i.e.,
instability scenario primarily due to bow-tie configurations). However, as already
discussed in “Algorithm rationale” section, we solved such problem relaxing the node
Fig. 2 Zachary Karate Club example. aThe Zachary Karate Club network. In (b-c) the bipartite graph showing
node-community affiliation (colored squares representing communities, yellow hexagons identifying nodes
belonging to more than one community). bClustering for φ=0.33, (c) Clustering for phi =0.66
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 9 of 23
single label constraint thus allowing for the identification of a stable configuration of
overlapping local communities.
ii) Merging: this step operates on a well-determined set of local communities on which
the PRECISIONMERGE procedure is applied iteratively. Since we explicitly impose the
community visit ordering the determinism of the solution is given by construction.
Property 2 (Compositionality)A
NGEL is easily parallelizable since the local com-
munity extraction can be applied locally on well defined subgraphs (i.e., ego-minus-ego
Given a graph G=(V,E)it is possible to instantiate ANGEL local community extraction
simultaneously on all the nodes uVand then apply the PRECISIONMERGE recursively
in order to reduce and compact the final overlapping partition:
Angel(G,φ) =PrecisionMerge
The underlying idea is to operate community merging only when all the local com-
munities have already been identified (i.e., LABELPROPAGATION is applied to all the
ego-minus-ego of the nodes uV– LP(EgoMinusEgo(u)) in Eq. 1– as shown in Fig. 3).
Moreover, this parallelization schema is assured to produce the same network partition
obtained by the original sequential approach due to the determinism property.
To evaluate the time complexity we proceed by decomposing ANGEL in its main com-
ponents. Given the pseudocode description provided in Algorithm 1 we can divide our
approach into the following sub-procedures:
i) Outer loop (lines 3-6): the algorithm cycles over all the nodes of the network to
extract the ego-minus-ego networks and identify local communities. This main loop
has thus complexity O(|V|).
ii) Local Communities extraction: the Label Propagation algorithm has complexity
O(n+m)(Raghavan et al. 2007), where
is the number of nodes and
is the
Fig. 3 ANGEL parallelization schema. The graph Gis decomposed in |V|ego-minus-ego network by a
dispatcher Dand distributed to nworkers {LP0,...,LPn} that extract local communities from them. At the end
of such parallel process, a collector Citeratively apply PRECISIONMERGE till obtaining the final overlapping
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 10 of 23
number of edges of the ego-minus-ego network. Let us assume that we are working
with a scale free network, whose degree distribution is pk=kα: in this scenario the
majority of the identified ego-minus-ego networks are composed by n<< |V|nodes
and m<< |E|edges, thus the average complexity of each iteration will be
iii) PRECISIONMERGE final cycle (lines 9-14): for each local community ANGEL evaluate
if it can be merged with one or more previously identified substructures. To efficiently
implement this task, we assume that once identified a community a new identifier is
generated and assigned to all the nodes within it. All the nodes will then have attached
multiple labels (one representing an identifier of a community the node belongs to).
Given a community
the PRECISIONMERGE function (Algorithm 2) leverage such
information to efficiently compute – for each community identifier
attached to the
nodes in
– the ratio of nodes in it that already belongs to
w.r.t. the size of
ratio is greater than (or equal to) a given threshold, the merge is applied and the node
label updated. This step can be performed with constant complexity employing an
hash-map, O(1). Considering the complete loop the overall cost is thus given by the
initial sorting of the communities by decreasing size, O(|C|log|C|)(where
is the
community set), and the evaluation of PRECISIONMERGE on each community in C,
O(|C|). Moreover, we can assume the number of iteration k<< |C|since at each
step the number of communities decreases: thus we can consider
as a constant
factor giving as final complexity, O(|C|log|C|)+O(|C|)=O(|C|log|C|).
Considered together such sub procedures gives us a final complexity of O(|V|(n+m)) +
O(|C|log|C|): considering a scale free network, for which we can reasonably expect
|V|>> (n+m)and |V|>|C|, the final complexity can be approximated as O(|V|).
ANGEL evaluation
Evaluating a community discovery approach is not an easy task. In this section, we pro-
pose a two-stage evaluation, focusing both on underlining ANGEL efficiency – in terms
of scalability and running time – as well as on its ability to retrieve ground-truth commu-
nities. As a first step, in “Competitors and datasets” section we identify the competitors
of our algorithm, approaches that share with it the same rationale. After that, in “Commu
nity resemblance” section, we briefly describe the quality function we adopt to compare
the partition produced by the selected algorithms and to assess their resemblance w.r.t.
ground-truth communities. Finally, we evaluate ANGEL and its competitors on two dif-
ferent community resemblance tasks: (i) identification of planted ground truth partition
in synthetically generated networks, “Synthetic benchmarks section, and (ii) identifica-
tion of semantic communities in real-world network datasets, “Evaluation on real world
Competitors and datasets
We de fined A NGEL as a two-phase bottom-up approach that leverage label propagation
to extract overlapping communities. To evaluate its performances, we compare it with
state-of-art competitors having a similar rationale2. In particular, our analysis includes:
2All the algorithms were executed on a Linux 4.4.0 machine with an Intel Core i7-5820 CPU @3.3GHzx16 and 32GB of
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 11 of 23
DEMON (Coscia et al. 2012;2014b) is an incremental and limited time complexity
algorithm for community discovery. It extracts ego networks, i.e., the set of nodes con-
nected to an ego node u, and identifies the real communities by adopting a democratic,
bottom-up merging approach of such structures.
PANDEMON (Amoretti et al. 2016) is a parallel implementation of DEMON designed to
increase its scalability and to reduce the computational complexity of its community
merging phase.
NODEPERCEPTION. In Soundarajan and Hopcroft (2015) the authors propose a gener-
alization of the DEMON approach: NODEPERCEPTION instantiate the local two-phase
schema by employing alternative community discovery approaches to Label Propagation
in the local community extraction phase. Thanks to such flexibility, NODEPERCEPTION
allows the final user to identify search for network partitions that optimize specific
quality functions.
SLPA. In Xie and Szymanski (2012) is introduced an overlapping hierarchical community
discovery algorithm designed for large-scale networks. SLPA leverages a label propaga-
tion strategy built upon dynamic interaction rules. The time complexity of SLPA scales
linearly with the number of edges in the network.
The former three (DEMON,PANDEMON and NODEPERCEPTION)movefromthesame
algorithmic schema of our approach. They all are node-centric algorithms (Rossetti et al.
2017b) that, moving from the analysis of ego-networks, generate overlapping partitions
following a non-deterministic approach and providing different computational complex-
ity. Conversely, the latter competitor, SLPA, represents a fast implementation of the label
propagation algorithm used by ANGEL,DEMON and PANDEMON to identify ego-network
local communities. Even though SLPA does not fall in the node-centric algorithmic fam-
ily, we decided to include it in our analysis since it can be seen as a baseline for all those
algorithms employing label propagation as the internal function.
Synthetic benchmarks. To ev a luat e how ANGEL behave under specific, controlled, set-
tings we tested it, along with its competitors, against synthetic networks having planted
ground truth communities generated through the LFR benchmark3(Lancichinetti et al.
2008). The networks described by LFR have well-known characteristics: among the oth-
ers, both their node degrees and community sizes follow a power law distributions.
Moreover, similar to the planted l-partition model(Condon and Karp 2001), LFR net-
work vertices share a predefined fraction of their links with other vertices of their cluster.
Finally, LFR allows the analyst to decide the average cluster density and size of the
generated graph. We generated multiple networks varying the following LFR parameters:
, the network size (from 100 to 100k nodes);
, the network density (from 0.1 to 0.4, steps of 0.1);
-μ, the mixing coefficient describing the average per-node ratio between the number
of edges to its communities and the number of edges with the rest of the network
(from 0.1 to 0.5, steps of 0.1).
3Code available at
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 12 of 23
Real world data. TTo understand how ANGEL and its competitors behave on real-
world data, we tested them against four network datasets having annotated ground-truth
community structure4. We analyzed the following datasets (Yang and Leskovec 2015),
whose synthetic statistics are briefly summarized in Table 1:
-emailEU. Network built upon email exchange data among members of a large
European research institution. The ground truth communities identify members’
-Amazon. Network built using the
Customers Who Bought This Item Also Bought
feature of the Amazon website. Each product category provided by Amazon defines
each ground-truth community.
-dblp. Co-authorship network where two authors are connected if they publish at
least one paper together. Publication venue, e.g., journal or conference, define
ground-truth communities.
-Youtube. Subgraph of the Youtube social network. User-defined groups identify
ground-truth communities.
Differently from synthetic benchmarks, where the planted communities respect specific
topological characteristics, real data annotation provides a semantic partition of network
nodes. Since none of the considered algorithms is parameter free in our analysis, we
instantiate each one of them multiple times performing a grid-search estimation of the
optimal parameter for each target network. Such parameter fitting strategy ensures that,
for each network, we compare the performances of the selected algorithms leveraging
their partitions that better approximate the ground truth ones.
Community resemblance
One way to asses the effectiveness of a CD algorithm is to compare how much the
communities it identifies can provide a good approximation of a given ground truth
partition. To quantify the degree of resemblance of two graph partitions we apply an
efficient methodology proposed in Rossetti et al. (2016). Given a community set X
produced by an algorithm and a ground truth community set Y,foreachcommunity
xXwe label its nodes with the ground truth community yYthey belong to. Then
we match community xwith the ground truth community with the highest number
of labels in the algorithm community. Such procedure produces (x,y)pairs having the
highest homophily between the node labels in xand all the ground truth communities.
The quality of the produced mappings is estimated in terms of precision and reacal:
Table 1 Datasets statistics
|V||E||C|CC d
emailEu 1,005 25,571 42 0.3994 7
Amazon 334,863 925,872 75,149 0.3967 44
dblp 317,080 1,049,866 13,477 0.6324 21
Youtube 1,134,890 2,987,624 8,385 0.0808 20
Number of nodes, edges, ground truth communities, average clustering coefficient and diameter for the analyzed datasets
4Datasets available at
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 13 of 23
Precision: identifies the percentage of nodes in xlabeled as y.Itisdefinedas:
|x|[0,1] (2)
Recall: identifies the percentage of nodes in ycovered by x.Itisdefinedas:
|y|[0,1]. (3)
Given a pair (x,y)the two scores describe the overlap of their members. A perfect match
is obtained when both precision and recall are equal to 1. Indeed, many-to-one map-
pings can occur: multiple communities in Xcanbeconnectedtoasinglegroundtruth
community in Y. This peculiarity allows the adoption of such methodology to evaluate
both algorithms producing crisp partitions as well as approaches producing overlapping
ones. In Rossetti et al. (2016), precision and recall are combined into their harmonic mean
obtaining the F1-measure, a concise quality score for the individual pairing:
F1=2precision recall
precision +recall .(4)
Given a network, the F1 score can be averaged among all the identified pairs in order to
summarize the overall correspondence between the algorithm community set and ground
truth community set. In the following, we will adopt a normalized version of the F1 score,
namely NF15, that mitigate the issues related to coverage and redundancy of communities
in assessing the final matching quality. In particular, defined as Yid the set of community
of Ymatched by community in X,wecandefineCoverage as:
Coverage =|Yid |
|Y|[0,1] (5)
and it identify the percentage of communities in Ythat are matched by at least an object
of C. Redundancy instead can be defined as:
Redundancy =|X|
Redundancy is minimized when no conflicting matches exist among the communities in
Xand the ones in Yid. Finally NF1 can be defined as:
Redundancy (0, 1] (7)
NF1 is maximized when: (i) the average F1 is maximal (perfect match), (ii) the community
in Xprovide a complete coverage for the ones in Yand (iii) the redundancy is mini-
mized (i.e., each community in Xis matched with a distinct community in Y). As shown
in Rossetti et al. (2016) it is possible to compute F1 (and thus NF1) paying a linear com-
plexity in the size of the community set X. The reduced complexity makes NF1 a suitable
alternative to the widely used NMI (Lancichinetti et al. 2008). Moreover, as discussed in
Lancichinetti et al. (2009)andMcDaidetal.(2011),NMIisnotstablewhilecomparing
overlapping partitions with non-overlapping ones while NF1 does not suffer such limita-
tion. Since all the compared algorithms produced overlapping partitions, NF1 represents
a reasonable resemblance function to adopt. Moreover, considering our analytical setup,
we avoid applying aggregate ranking solutions (as proposed in Jebabli et al. (2018)and
(Dao et al. 2018)) that are designed to obtain a more comprehensive view of cluster-
ing properties. Evaluation approaches belonging to such a family generate an aggregate
5Code available at
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 14 of 23
ranking able to summarize the behaviors of alternative CD solutions once that multiple
clustering fitness/comparison scores are available (see, for instance, Orman et al. (2012)).
However, considering that all the compared algorithms are intended to operate under the
same rationale, we feel that focusing on NF1 provides us enough information to draw a
few considerations on the obtained clusterings.
Experimental results
Synthetic benchmarks
In Fig. 4we report the execution time and NF1 score for the compared CD approaches.
Our experiments show that our approach is able sensibly to improve the running times of
its competitors while increasing the network size. In particular, it is worth noticing that
Fig. 4a reports execution times on a log scale: considering the average runtime of ANGEL
on the generated 100k nodes graphs it registers a speedup of an order of magnitude w.r.t.
itscompetitors.InFig.4b-c the NF1 score is used to compare the adherence of the par-
titions identified by the selected CD algorithms to the ground truth ones: we omitted
NODEPERCEPTIONs results since their overall NF1 were always lower than 0.4. In partic-
ular Fig. 4b compare the average NF1 scores obtained by each algorithm on different sized
LFR graphs. To compute the NF1 mean value for the pair <algorithm,network_size >we
considered the results provided by the optimal parameter configuration w.r.t. each net-
work size instantiation (e.g., varying graph density and mixing coefficient). Among the
compared methods ANGEL is always able to reach the highest scores, often producing the
perfect match for the planted communities. Figure 4c underline the impact of network
mixing coefficient on the quality of extracted communities once fixed the network size.
We ob ser ve that A NGEL and SLPA can assure relatively stable performances while varying
Evaluation on real world data
Tabl e 2shows the running times – expressed in seconds – of the compared CD
approaches when applied to the selected networks. As already underlined in the synthetic
scenario, ANGEL outperforms its competitors, often achieving execution times of one or
more orders of magnitude less.
Differently from the synthetic scenario, when it comes to assessing community resem-
blance – quantitative values in Table 3– we observe a relatively low quality for all
the partitions produced by the compared algorithms. Indeed, such results are somehow
Fig. 4 Synthetic Benchmarks. aRunning time of the compared algorithms w.r.t. network size (number of
nodes); bAvg. community resemblance score per network size; cCommunity resemblance varying LFR
mixing coefficient (number of nodes 106). In (b-c) the NF1 scores for NODEPERCEPTION are omitted due to
their low values (NF10.2)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 15 of 23
Table 2 Running times
Angel Demon PanDemon NodePerception SLPA
emailEu 3.53 4.72 2.34 9.91 2.42
Amazon 88.49 16862.61 3032.63 256.09 504.61
dblp 115.44 24273.36 1059.54 382.43 321.46
Youtube 2209.20 8362.28 4076.98 11533.74 2860.01
The execution times reported are expressed in seconds and do not include network loading and results serialization on file.
PANDEMON has been executed on 16 cores
expected. Conversely, from the synthetic benchmark where the planted communities
were designed to follow specific topological characteristics, the semantic annotation pro-
vided for the analyzed real-world network do not necessarily reflect structural properties
(Hric et al. 2014). Such decoupling makes difficult, if not impossible, for CD algorithms
that do not leverage semantic information to capture the same partition identified by the
ground truth. However, even in this more complex scenario, ANGEL communities are the
ones able to better approximate the provided ground truth node partitions. In order to
provide a statistical significance bound to our experiment on real data we also applyed a
Friedman test (Friedman 1937) with Li post-hoc evaluation (Li 2008) on the evaluation
proposed in Table 3. The test was rejected for the NF1 scores with a p-value of 0.05, thus
implying that the compared methods do actually behave differently when tested on mul-
tiple datasets. Moreover, the post-hoc underlined that Angel significantly outperforms
NODEPERCEPTION under the same confidence interval, and all the others when p-value
is imposed equalt to 0.1.
ANGEL on dynamic networks
In this section, we propose an extension of ANGEL tailored to extract communities from
dynamic network topologies so to observe their evolution as time goes by. From a mod-
elling point of view, in the following sections, we represent a dynamic network by using
snapshot graphs:
Definition 3 (Snapshot Graph) Let G be an attributed graph G =(V,E,T),whereV
is a set of nodes, E a set of edges and T ={0, 1 ...,n}an ordered set of labels (associated
to both nodes and edges) identifying different timestamps. Given a label i T, we call
Gi=(Vi,Ei)the graph induced from G composed by the nodes, edges whose labels is i. A
Snapshot Graph is defined as the set G={G0,G1...Gn}composed by n consecutive, non
temporal overlapping, partition of G such as G =∪
Using such temporal discretization in Instant-Optimal dynamic community discovery
we extend ANGEL to handle dynamic networks and briefly frame the resulting approach
Table 3 Community resemblance
Angel Demon PanDemon NodePerception SLPA
emailEu 0.51 0.20 0.04 0.12 0.23
Amazon 0.17 0.10 0.12 0.05 0.09
dblp 0.52 0.32 0.52 0.02 0.27
Youtube 0.19 0.08 0.08 0.04 0.18
NF1 scores achieved by the compared algorithms on the real world datasets. For each model, we report the score achieved by its
optimal parameter settings
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 16 of 23
within a specific subclass of DCD methodologies: Instant Optimal,Identif y&Match algo-
rithms. Finally, in Complexity we discuss the computational complexity of the proposed
Instant-Optimal dynamic community discovery
As previously discussed, ANGEL efficiently address the classical formulation of the over-
lapping community discovery problem: however, per se it is not designed to take into
account the challenges that evolving network topologies generate. The natural way to
proceed, to enable ANGEL to dynamic community analysis, is to extend it by applying an
algorithmic schema known as Identify&Match (or, equivalently, “Two-steps"(Alhajj and
Jon 2014)). Such schema characterizes a vast majority of DCD approaches originating
from the extension of static methods to dynamic topologies. As suggested by its name,
such strategy describes a two steps process:
: detect static communities on each step of evolution;
: align the communities identified at step
with the ones at step t1.
The main advantage of two-steps solutions is that they allow reusing static CD techniques,
avoiding the definition of novel, often context dependent, methodologies. Moreover, one
of the reasons for the abundance of methods belonging to such family lies in the fact
that the matching step can be derived from existing literature, since set matching is a
widely studied problem. Moreover, Identify&Match allows to easily describe parallelizable
analytical workflows. As discussed in (Rossetti and Cazabet 2018), “Two-step" approaches
represent a specialization of a more general class of algorithms, called Instant-Optimal
CD. Indeed, matching the communities found at different stages of network evolution
might involve comparing several sets of temporally disjoint network partitions: however,
Instant Optimal CD approaches assume that the partition identified at tis optimal, w.r.t.
the topology of the network at t. DCD solutions falling in this class are, by definition,
non-temporally smoothed and represent the best choice when the final goal is to provide
communities which are as good as possible at each step of the evolution of the network.
Algorithm 3 details the pseudocode of the dynamic extension of ANGEL.Thealgorithm
required inputs are (i) a set of snapshot graphs, and (ii) the φthreshold (as requested by
ANGEL). To avoid having a third parameter (e.g., a similarity threshold for the matching
phase), we don’t make use of the Jaccard similarity – a widely adopted strategy to address
this kind of approaches – while aligning community sets extracted from consecutive net-
work snapshots. Conversely, we adopted a matching criterion similar to the one used
to merge communities in the second phase of ANGEL, thus providing a coherent merg-
ing/matching strategy. We assume that each node at time tcarries three sets of labels: i)
the identifiers of the communities it currently belongs to; ii) the identifiers of the commu-
nities it was part of at t1, and; iii) the identifiers of the communities it will be associated
to at t+1. Given two community sets – i.e., the ones at time t1andt–construct-
ing the requested labelling has linear complexity in the number of the nodes. Once the
nodes belonging to temporally adjacent partitions are labelled, the following matching
procedure is performed:
i) firstly, each community identified in Gt1is matched with the ones in Gtthat
maximize the precision score;
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 17 of 23
Input:G={G0,G1,...Gn}, a snapshot graph; φ, the merging threshold.
Output:L, a set of community lifecycles.
2for GiGdo
3CANGEL(Gi,φ); // Community Extraction
4LPrecision(Ci,Ci1); // Merge Detection
5LLPrecision(Ci1,Ci); // Split Detection
6return L
ii) secondly, the same criterion is used to match each community in Gtwith the more
similar ones in Gt1.
Indeed, the precision score (as defined in Equation 2) is not symmetric, thus performing
the matching in both directions makes possible the identification of different evolutive
patterns involving the observed communities.
Since the algorithm core is based on the ANGEL one, it is clear that its community identifi-
cation step has complexity O(|V|). Assuming no parallelism, an equal number of nodes in
each graph snapshot, and ksnapshots composing the network evolution, we get O(k|V|).
Each merging phase, after the labelling step having complexity O(|v|),hascomplexity
O(|C|),whereCis the identified community set (as discussed in (Rossetti et al. 2016)):
thus assuming kcommunity sets of approximately the same size, we get O((k1)|C|).
Since we can assume that k<< |C|<< |V|the overall complexity is O(|V|).
Dynamic ANGEL evaluation
Symmetrically to what we did in “ANGEL evaluation” section, here we evaluate our DCD
approach both on synthetic and real-world data. Moving to a dynamic scenario two differ-
ent quality functions should be taken into consideration while evaluating a CD approach:
community resemblance and matching effectiveness. To assess the former, coherently to
what we did for ANGEL, we will proceed leveraging the NF1 score; to evaluate the latter,
on the other hand, we will analyze to what extent the community events identified by our
approach are in line with the ones annotated in the data. Since community events repre-
sent a peculiarity of time evolving contexts in Community events and life-cycle we firstly
introduce a high-level discussion of what is usually called Community Lifecycle. Finally,
in Synthetic benchmarks and Evaluation on real world data we describe and discuss the
results obtained on synthetic benchmarks as well as on real data. Conversely, from what
previously done for ANGEL we will not compare the obtained results by its extension with
the ones produced by others approaches following the same rationale: in this scenario, our
focus will be on the analysis of the discovered patterns and the evaluation of the proposed
matching strategy.
Moreover, so far most the two-step approaches proposed in literature leverage non-
overlapping static CD algorithms whose definition is not aligned to the ANGEL node-
centric bottom-up one. Finally, a comparison with DCD extensions of the competitors
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 18 of 23
identified in “Competitors and datasets” section can only focus on the matching effective-
ness since the resemblance has already been evaluated in the static scenario.
Community events and life-cycle
Several works on evolutionary community discovery focus on the analysis of the events
which regulates community life-cycles – birth, merge, split, continue and death of
communities. Such events are commonly described as follows:
Birth (B): a community born at time
if there are no network substructures at t1
that can be matched with it;
Merge: two or more communities at time
merge iff they are matched to the same
network substructure at t+1;
Split (S): a community at
splits if it is matched to multiple network substructures at
Continue (C): a community at
remains the same at t+1;
Death (D): a community dies at
if it is not matched with any network substructure
at t+1.
Indeed, rarely a generic DCD algorithm is able to track all the community events intro-
duced. However, some events can be easily reconstructed as a post-process analytical step
from the results of every DCD approach.
The main open issue that affects the evaluation of community life-cycles lies in the so-
called Theseus’s ship paradox. During his adventurous journey, the Greek hero Theseus
had to replace his ship piece by piece due to the adversities he had to face on the sea.
Once returned home, Theseus’s ship is placed in a museum as the iconic memory of the
hero legacy. In its essence, the ship was still the same as the one Theseus used when he
left Troezen; however, none of its original components were still part of it. Analyzing
the evolution of communities leads to a similar dilemma: an evolving community can be
considered as such if during his history most (if not all) the nodes that compose it change?
Indeed multiple answers can be provided to such question: in our analysis, we focus
our attention on community event trends through time, without directly constructing
individual community life-cycle.
Experimental results
Synthetic benchmarks
SimilarlytowhatwehavedoneforANGEL, we evaluate its extension against synthetic
dynamic networks having planted ground truth communities. So far few benchmarks
specifically designed for the DCD problem have been proposed: among them, we recall
Greene’s (Greene et al. 2010;Granelletal.2015). It composes of four dynamic networks –
observed for five snapshots – each one describing a specific kind of community evolution
event. In particular, Greene benchmarks aim to evaluate the ability of a DCD approach in
identifying the following mesoscale topology perturbations:
. In the corresponding network, 10% of communities are removed
randomly from each snapshot to its successor;
. To examine the effect of rapid expansions and
contractions of communities, in this network in each snapshot 40 randomly selected
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 19 of 23
communities, absorb new nodes or lose their former members by 25% of the previous
. At each snapshot, 40 new clusters are created by the nodes which
have left their former communities. Furthermore, 40 existing communities are
removed randomly;
. In this last setting, during each snapshot, 40 instances of
existing clusters are merged two by two, and similarly, 40 communities split into two
new communities.
Such benchmarks can be used to evaluate both community resemblance and event match-
ing since they explicitly define both ground truth network partitions for each snapshot as
well as the events relating communities of consecutive snapshots.
Indeed, Greene’s benchmark is not the only one designed to test DCD approaches. In
the last decade, several dynamic graph generators have been proposed to address such
issues (Rossetti 2017;Bazzietal.2016;Linetal.2008), often extending/revising the
LFR approach. In the following, we evaluate the proposed algorithm employing only the
Greene’s synthetic networks corpus since it can be seen as a de-facto standard, being the
most widely adopted benchmark to test DCD algorithms. Figure 5show the results that
the dynamic extension of ANGEL achieves when tested against Greene’s benchmark. As a
first result, Fig. 5a underlines the trends for community coverage in all the described sce-
narios. The reported trend lines underline that ANGEL can identify most of the ground
truth community of each snapshot. The partition coverage – i.e., the number of ground-
truth communities matched by the ones extracted by our approach – assumes values in
[0.94, 0.98]. Moreover, in Fig. 5b, we report the match quality, , of the communities
involved in merge/split, birth/death, Expand and Hide events. is a function that relates
the set of discovered events, A, to the ground truth ones, B:
The quality function assumes values in [-|A|, 1]: when maximized, it implies a per-
fect match of the two event sets when minimized it identify completely disjointed event
sets, values in between capturing both the presence of partial matches and wrong ones.
Our results underline that the proposed matching strategy can identify the planted events
that – with the sole exception of Birth/Death in the first two matching steps – are per-
fectly reconstructed. Finally, in Fig. 5b we report the overall community quality score
computed using the previously introduced NF1 score. Moreover, resemblance scores of
ANGEL communities w.r.t. the considered benchmark assume values in [0.91, 0.95], thus
Fig. 5 Greene Benchmarks. aCommunity coverage per snapshot and graph sequence; bPercentage of
communities matched per snapshots and graph sequence; cNF1 score per snapshot and graph sequence
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 20 of 23
highlighting the ability of the proposed method to extract partitions that adhere to the
original ground-truth.
Evaluation on real world data
Temporally evolving networks are often used to describe real-world phenomena.
Dynamic network data rarely come with annotated evolving ground-truth communities;
however, we decided to test our approach on real-world datasets to provide a charac-
terization of the partitions our approach identifies. To do so, we analyzed the following
datasets6, whose synthetic statistics are reported in Table 4:
-Enron. The Enron email network consists of emails sent among the employees of
Enron Corporation between the beginning of 1999 and 2002. Graph snapshots are
taken monthly.
-FB-wosn. Dynamic network built upon the wall posts from the Facebook New
Orleans networks. Each edge represents a post to other users wall, and each node
represents a Facebook user. The time interval between snapshots is selected one
-Weib o. This dataset is obtained from the 2012 WISE Challenge7. It is built upon the
logs of the popular Chinese micro-blog service WEIBO8. Edges represent mentions
made by users in short messages. We selected a single year, 2011, and used an
observation window of one month to build the snapshots.
As stated, differently from the real data analyzed in the static scenario, we do not have
any ground-truth annotation for the communities nor for the events they experience. To
provide some insights on the results obtained, we report in Fig. 6the trends of commu-
nity events through time. Communities can be seen as the bricks of a complex network
structure, the way they interact (both merging and splitting), form and dissolve, describe
the pulse of a dynamic system. A leitmotif shared by all the networks analyzed regards
the continue event being the least represented. As expected, and somehow postulated by
the Theseus’s ship paradox, it is rare that a community remains stable (i.e., composed by
the same exact set of nodes) as time goes by, especially in social interaction networks like
the ones we considered. Apart from this common result, the three networks seem to be
characterized by different event-patterns. In Fig. 6a we can observe overall stationarity
of the Weibo service: all the trends seems to maintain a predictable increasing pattern,
interrupted only by a sudden change in the number of communities during August 2011.
The FB-wosn data tells a completely different story. Figure 6b describe a progressive,
steady, increase in the number of communities, underlying that, during each observa-
tion, the relative importance of events remains almost constant. In this scenario, the birth
Table 4 Datasets statistics
Enron 87,110 298,691 48 0.1196
FB-wosn 61,096 614,797 36 0.1886
Weibo 8,335,605 49,595,797 12 0.0146
The number of nodes, edges, snapshots, average clustering coefficient for the analyzed datasets. Network related statistics are
computed on flattened graphs
6Datasets available at:
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 21 of 23
Fig. 6 Event Trends. For each dynamic network are reported the trends (in absolute numbers) of community
events. Moreover, the trends of identified communities as time goes by is also shown
and death events are the most frequent ones: communities change quickly and radically
from snapshot to snapshot, allowing less gradual merge/splits than in Weibo. Finally,
Fig. 6c describes a very peculiar trend emerging from the emails exchanged between
Enron employees. As a unique among the observed datasets, in Enron, the merge and split
events are constantly the most prominent ones. Conversely, from FB-wosn and Weibo,
online social platforms whose users can deeply vary among consecutive observation
periods, Enron nodes are particularly stable. Node stability, as well as routinary work-
related communications, reduce the likelihood of identifying the formation of completely
novel communities. Indeed, changing the snapshot temporal granularity could affect such
results: for instance, moving from a monthly to a daily temporal discretization in Enron
generates an increase of Death/Birth events during weekends, reflecting reduced email
In this paper, we introduced ANGEL, a node-centric approach to overlapping community
discovery. ANGEL was designed with the aim of lowering the computational complexity
of existing approaches while ensuring the identification of high-quality partitions. Exper-
imental results, both on synthetic and real-world networks, highlight the efficiency and
effectiveness of the proposed approach, underlying its ability to outperform its direct
Since such topologies usually evolve as time goes by, we also introduced an extension
of ANGEL, a dynamic community discovery approach that allows for tracking and analyz-
ing dynamic communities life-cycles. We evaluate our DCD approach both quantitatively
on well known synthetic benchmarks and qualitatively on real data. As future work, we
plan to apply the proposed algorithms as support for network analytical tasks and, dur-
ing the process, to implement and evaluate parallel and distributed ANGEL variants able
to scale up to big data. Moreover, we also plan to approach the “Theseus’s ship paradox"
with the aim of proposing novel post-processing heuristics able to consolidate community
life-cycles (independently from the DCD approach used to generate them) thus avoiding
instability phenomena and ambivalent interpretations.
CD: Community discovery; DCD: Dynamic community discovery; LP: Label propagation; NF1: Normalized F1
Not applicable.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 22 of 23
Authors’ contributions
G.R. designed the algorithms, performed the experiments, validated the results and wrote the paper. The author(s) read
and approved the final manuscript.
This work is supported by the scheme ’INFRAIA-01-2018-2019: Research and Innovation action’, Grant Agreement n.
871042 ’SoBigData++: European Integrated Infrastructure for Social Mining and Big Data Analytics’
Availability of data and materials
The developed algorithms are released under the BSD-license 2 clause, their implementation is available at https:// and within the CDlib python library
datasets generated and/or analysed during the current study are available in the Netorkrepository repository (http://
Competing interests
The authors declare that they have no competing interests.
Received: 12 March 2020 Accepted: 27 May 2020
Alhajj R, Jon R (2014) Encyclopedia of social network analysis and mining. Springer Publishing Company, Incorporated
Amoretti M, Ferrari A, Fornacciari P, Mordonini M, Rosi F, Tomaiuolo M (2016) Local-first algorithms for community
detection. In: KDWeb
Arnaboldi V, Conti M, Passarella A, Dunbar RIM (2017) Online social networks and information diffusion: The role of ego
networks. OSNEM 1:44–55.
Bazzi M, Jeub LG, Arenas A, Howison SD, Porter MA (2016) Generative benchmark models for mesoscale structure in
multilayer networks. arXiv preprint arXiv:1608.06196
Buzun N, Korshunov A, Avanesov V, Filonenko I, Kozlov I, Turdakov D, Kim H (2014) Egolp: Fast and distributed community
detection in billion-node social networks. In: ICDMW. IEEE.
Casteigts A, Flocchini P, Quattrociocchi W, Santoro N (2012) Time-varying graphs and dynamic networks. Int J Parallel
Emergent Distrib Syst 27(5):387–408
Cazabet R, Amblard F, Hanachi C (2010) Detection of overlapping communities in dynamical social networks. In: 2nd
International Conference on Social Computing (SocialCom). IEEE. pp 309–314.
Cazabet R, Rossetti G, Amblard F (2017) Dynamic community detection. In: Encyclopedia of Social Network Analysis and
Mining. Springer. pp 1–10. 4614-6170-8_383
Condon A, Karp RM (2001) Algorithms for graph partitioning on the planted partition model. Random Struct Algorithm
Coscia M, Giannotti F, Pedreschi D (2011) A classification for community discovery methods in complex networks. Stat
Anal Data Min.
Coscia M, Rossetti G, Giannotti F, Pedreschi D (2012) Demon: a local-first discovery method for overlapping communities.
In: International Conference on Knowledge Discovery and Data Mining. ACM. pp 615–623.
Coscia M, Rossetti G, Giannotti F, Pedreschi D (2014) Uncovering hierarchical and overlapping communities with a
local-first approach. ACM Trans Knowl Discov Data (TKDD) 9(1):6
Coscia M, Rossetti G, Giannotti F, Pedreschi D (2014) Uncovering hierarchical and overlapping communities with a
local-first approach. Trans Knowl Discov Data 9(1):6
Dao V-L, Bothorel C, Lenca P (2018) Community structure: A comparative evaluation of community detection methods.
arXiv preprint arXiv:1812.06598
Epasto A, Lattanzi S, Paes Leme R (2017) Ego-splitting framework: from non-overlapping to overlapping clusters. In:
SIGKDD. ACM. pp 145–154.
Fortunato S (2010) Community detection in graphs. Phys Rep 486(3).
Fortunato S, Hric D (2016) Community detection in networks: A user guide. Phys Rep 659:1–44
Friedman M (1937) The use of ranks to avoid the assumption of normality implicit in the analysis of variance. J Am Stat
Assoc 32.
Granell C, Darst RK, Arenas A, Fortunato S, Gómez S (2015) Benchmark model to assess community structure in evolving
networks. Phys Rev E 92(1):012805
Greene D, Doyle D, Cunningham P (2010) Tracking the evolution of communities in dynamic social networks. In:
Advances in Social Networks Analysis and Mining (ASONAM), 2010 International Conference On. IEEE. pp 176–183.
Himmel A-S, Molter H, Niedermeier R, Sorge M (2016) Enumerating maximal cliques in temporal graphs. In: IEEE/ACM
International Conference on Advances in Social Networks Analysis and Mining (ASONAM). IEEE. pp 337–344. https://
Holme P, Saramäki J (2012) Temporal networks. Phys Rep 519(3):97–125
Hric D, Darst RK, Fortunato S (2014) Community detection in networks: Structural communities versus ground truth. Phys
Rev E 90(6):062805
Jebabli M, Cherifi H, Cherifi C, Hamouda A (2018) Community detection algorithm evaluation with ground-truth data.
Physica A Stat Mech Appl 492:651–706
Kumpula JM, Kivelä M, Kaski K, Saramäki J (2008) Sequential algorithm for fast clique percolation. Phys Rev E 78(2):026109
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Rossetti Applied Network Science (2020) 5:26 Page 23 of 23
Lancichinetti A, Fortunato S, Kertész J (2009) Detecting the overlapping and hierarchical community structure in complex
networks. New J Phys.
Lancichinetti A, Fortunato S, Radicchi F (2008) Benchmark graphs for testing community detection algorithms. Phys Rev E
Li JD (2008) A two-step rejection procedure for testing multiple hypotheses. J Stat Plan Infer 138(6):1521–1527
Lin Y-R, Chi Y, Zhu S, Sundaram H, Tseng BL (2008) Facetnet: a framework for analyzing communities and their evolutions
in dynamic networks. In: Proceedings of the 17th International Conference on World Wide Web (WWW). ACM.
pp 685–694.
Malliaros FD, Vazirgiannis M (2013) Clustering and community detection in directed networks: A survey. Phys Rep
Matias C, Miele V (2016) Statistical clustering of temporal networks through a dynamic stochastic block model. J R Stat
Soc Ser B (Stat Methodol).
McDaid AF, Derek G, Neil H (2011) Normalized mutual information to evaluate overlapping community finding
algorithms. arXiv preprint arXiv:1110.2515
Milli L, Monreale A, Rossetti G, Pedreschi D, Giannotti F, Sebastiani F (2015) Quantification in social networks. In: 2015 IEEE
International Conference on Data Science and Advanced Analytics (DSAA) Paris. pp 1–10.
Moradi F, Olovsson T, Tsigas P (2014) A local seed selection algorithm for overlapping community detection. In: ASONAM.
Morini M, Flandrin P, Fleury E, Venturini T, Jensen P (2017) Revealing evolutions in dynamical networks. arXiv preprint
Mucha PJ, Richardson T, Macon K, Porter MA, Onnela J-P (2010) Community structure in time-dependent, multiscale, and
multiplex networks. Science 328(5980):876–878
Orman GK, Labatut V, Cherifi H (2012) Comparative evaluation of community detection algorithms: a topological
approach. J Stat Mech Theory Exp 2012(08):08001
Palla G, Barabási A-L, Vicsek T (2007) Quantifying social group evolution. Nature 446(7136):664–667
Raghavan UN, Albert R, Kumara S (2007) Near linear time algorithm to detect community structures in large-scale
networks. Phys Rev E 76(3).
Rossetti G (2017) Rdyn: Graph benchmark handling community dynamics. J Complex Netw.
Rossetti G, Cazabet R (2018) Community discovery in dynamic networks: A survey. ACM Comput Surv (CSUR) 51(2):35
Rossetti G, Luca P, Salvatore R (2016) A novel approach to evaluate community detection algorithms on ground
truthComplex Networks VII. Springer, Cham. pp 133–144
Rossetti G, Milli L, Cazabet R (2019) Cdlib: a python library to extract, compare and evaluate communities from complex
networks. Appl Netw Sci. 0165-9
Rossetti G, Pappalardo L, Kikas R, Pedreschi D, Giannotti F, Dumas M (2015) Community-centric analysis of user
engagement in skype social network. In: ASONAM. ACM.
Rossetti G, Pappalardo L, Kikas R, Pedreschi D, Giannotti F, Dumas M (2016) Homophilic network decomposition: a
community-centric analysis of online social services. Soc Netw Anal Min J 6(103).
016-0411- 4
Rossetti G, Pappalardo L, Pedreschi D, Giannotti F (2017a) Tiles: an online algorithm for community discovery in dynamic
social networks. Mach Learn 106(8):1213–1241
Rossetti G, Pappalardo L, Rinzivillo S (2016) A novel approach to evaluate community detection algorithms on ground
truth. In: Complex Networks.
Rossetti G, Pedreschi D, Giannotti F (2017b) Node-centric community discovery: From static to dynamic social network
analysis. OSNEM 3:32–48
Soundarajan S, Hopcroft JE (2015) Use of local group information to identify communities in networks. Trans Knowl
Discov Data.
Takaffoli M, Sangi F, Fagnan J, Zaïane OR (2011) Modec-modeling and detecting evolutions of communities. In: 5th
International Conference on Weblogs and Social Media (ICWSM). AAAI. pp 30–41
Viard T, Latapy M, Magnien C (2016) Computing maximal cliques in link streams. Theor Comput Sci 609:245–252
Whang JJ, Gleich DF, Dhillon IS (2016) Overlapping community detection using neighborhood-inflated seed expansion.
IEEE Trans Knowl Data Eng 28(5):1272–1284
Xie J, Kelley S, Szymanski BK (2013) Overlapping community detection in networks: The state-of-the-art and comparative
study. Comput Surv.
Xie J, Szymanski BK (2012) Towards linear time overlapping community detection in social networks. In: PAKDD. https:// 642-30220-6_3
Yang J, Leskovec J (2015) Defining and evaluating network communities based on ground-truth. Knowl Inf Syst
Zakrzewska A, Bader DA (2015) A dynamic algorithm for local community detection in graphs. In: Proceedings of the
2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM). ACM.
pp 559–564.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
... We also note that all the clustering methods we developed produce disjoint clusters, yet publications may be expected to be members of more than one research community. Some clustering methods have been developed that can produce overlapping clusters (Rossetti, 2020), and exploring this approach in the context of large citation networks is likely to provide additional insights. Methods that can combine information from multiple clustering methods could lead to better insight into community structure, and so principled development of ensemble methods (a standard approach in machine learning) is another direction for future research. ...
... One of the most intriguing directions for future research is the life cycle of these research communities, both in terms of how ideas and questions being focused on by the research community as a whole change over time and how authors move between different communities over time. This is challenging to study because the network evolves through growth each year and community finding in dynamic graphs is a promising direction (Rossetti, 2020). ...
Full-text available
Clustering and community detection in networks are of broad interest and have been the subject of extensive research that spans several fields. We are interested in the relatively narrow question of detecting communities of scientific publications that are linked by citations. These publication communities can be used to identify scientists with shared interests who form communities of researchers. Building on the well-known k-core algorithm, we have developed a modular pipeline to find publication communities. We compare our approach to communities discovered by the widely used Leiden algorithm for community finding. Using a quantitative and qualitative approach, we evaluate community finding results on a citation network consisting of over 14 million publications relevant to the field of extracellular vesicles.
... In [26], the authors propose a greedy algorithm that starts from a heuristic initial solution and then it iteratively improves the solution by applying two transformation steps, deconstruction and reconstruction, in order to avoid local solutions and make a wider exploration of the search space. There is also interesting research on topics directly related to the community detection task, such as optimization for very large networks [27], or specific evaluation metrics for the task [28]. Finally, recent approaches like [29]- [31] propose the use of embeddings to characterize graph communities, analyzing the relationship among graph embeddings, community detection and node embeddings. ...
Full-text available
The detection of communities in social networks is a task with multiple applications both in research and in sectors such as marketing and politics among others. In this paper, we address the task of detecting on-line communities of Twitter users for a given domain. Our main contribution consists in modelling the community detection problem as a biclustering task. We have performed the experimentation with data from the political domain, a very dynamic area with a large number of interested users and a high availability of tweets. We have evaluated our proposal using both extrinsic and intrinsic methods, reaching very good results in both cases. We use the silhouette coefficient as intrinsic metric for clustering evaluation, and a classification task of political leanings of Twitter users as extrinsic evaluation. One of the most interesting conclusions of our experiments is the quality, from the point of view of predictive capacity in the classification task, of the communities identified with the proposed method. The information provided by communities detected through “follow” relationships has a predictive capacity comparable to that of the contents of tweets written by users. The results also show how detected communities can give insights about future events related to these communities that arise around social networks.
Detecting and analyzing community structure is a challenging topic in dynamic social network analysis. Although the number of methods in this area is on the rise, there are only a few algorithms that can discover meaningful communities based on different aspects of social networks. Indeed, social networks contain various information sources that can be used to analyze them. The most important part of this information is related to users’ topics of interest (content information) and users’ interactions (structure information). One promising solution to discover meaningful communities is to combine these two concepts. Based on this, we introduce ACSIMCD, a 2-phase framework for discovering and updating community structure without recomputing them from scratch at each snapshot. This article mainly includes two parts. In the first part, a static community detection algorithm which is called Content and Structure Information based Method for Community Detection (CSIMCD for short) is proposed to discover the initial community structure. The CSIMCD uses a hybrid approach founded on statistical and semantic measures to extract the users’ topics of interest. Accordingly, the original network is divided into several clusters (topical clusters) so that each one represents a distinct topic, then by performing a link analysis on each topical cluster, the communities are detected. In the second part, we propose ACSIMCD (Adaptive CSIMCD), an adaptive method for detecting and updating community structure in dynamic social networks. More precisely, the ACSIMCD explores the topics of interest of each changed node to identify the topical cluster it belongs to. After that, we update the community structure in this topical cluster, and we keep others as they are. We compare the ACSIMCD model with algorithms from different approaches including content-based methods on real-world networks. The experimental results showed that ACSIMCD produces a community structure of high quality from the perspective of links and interests compared with the classical methods, and that it is able to process network changes effectively in a reasonable time scale.
Full-text available
Abstract Community Discovery is among the most studied problems in complex network analysis. During the last decade, many algorithms have been proposed to address such task; however, only a few of them have been integrated into a common framework, making it hard to use and compare different solutions. To support developers, researchers and practitioners, in this paper we introduce a python library - namely CDlib - designed to serve this need. The aim of CDlib is to allow easy and standardized access to a wide variety of network clustering algorithms, to evaluate and compare the results they provide, and to visualize them. It notably provides the largest available collection of community detection implementations, with a total of 39 algorithms.
Full-text available
The past few years have seen intensive research efforts carried out in some apparently unrelated areas of dynamic systems-delay-tolerant networks, opportunistic-mobility networks, social networks-obtaining closely related insights. Indeed, the concepts discovered in these investigations can be viewed as parts of the same conceptual universe; and the formal models proposed so far to express some specific concepts are components of a larger formal description of this universe. The main contribution of this paper is to integrate the vast collection of concepts, formalisms, and results found in the literature into a unified framework, which we call TVG (for time-varying graphs). Using this framework, it is possible to express directly in the same formalism not only the concepts common to all those different areas, but also those specific to each. Based on this definitional work, employing both existing results and original observations, we present a hierarchical classification of TVGs; each class corresponds to a significant property examined in the distributed computing literature. We then examine how TVGs can be used to study the evolution of network properties, and propose different techniques, depending on whether the indicators for these properties are a-temporal (as in the majority of existing studies) or temporal. Finally, we briefly discuss the introduction of randomness in TVGs.
Full-text available
The past few years have seen intensive research efforts carried out in some apparently unrelated areas of dynamic systems-delay-tolerant networks, opportunistic-mobility networks and social networks-obtaining closely related insights. Indeed, the concepts discovered in these investigations can be viewed as parts of the same conceptual universe, and the formal models proposed so far to express some specific concepts are the components of a larger formal description of this universe. The main contribution of this paper is to integrate the vast collection of concepts, formalisms and results found in the literature into a unified framework, which we call time-varying graphs (TVGs). Using this framework, it is possible to express directly in the same formalism not only the concepts common to all those different areas, but also those specific to each. Based on this definitional work, employing both existing results and original observations, we present a hierarchical classification of TVGs; each class corresponds to a significant property examined in the distributed computing literature. We then examine how TVGs can be used to study the evolution of network properties, and propose different techniques, depending on whether the indicators for these properties are atemporal (as in the majority of existing studies) or temporal. Finally, we briefly discuss the introduction of randomness in TVGs.
Full-text available
Community structure is of paramount importance for the understanding of complex networks. Consequently, there is a tremendous effort in order to develop efficient community detection algorithms. Unfortunately, the issue of a fair assessment of these algorithms is a thriving open question. If the ground-truth community structure is available, various clustering-based metrics are used in order to compare it versus the one discovered by these algorithms. However, these metrics defined at the node level are fairly insensitive to the variation of the overall community structure. To overcome these limitations, we propose to exploit the topological features of the 'community graphs' (where the nodes are the communities and the links represent their interactions) in order to evaluate the algorithms. To illustrate our methodology, we conduct a comprehensive analysis of overlapping community detection algorithms using a set of real-world networks with known a priori community structure. Results provide a better perception of their relative performance as compared to classical metrics. Moreover, they show that more emphasis should be put on the topology of the community structure. We also investigate the relationship between the topological properties of the community structure and the alternative evaluation measures (quality metrics and clustering metrics). It appears clearly that they present different views of the community structure and that they must be combined in order to evaluate the effectiveness of community detection algorithms.
Full-text available
Several research studies have shown that complex networks modeling real-world phenomena are characterized by striking properties: (i) they are organized according to community structure, and (ii) their structure evolves with time. Many researchers have worked on methods that can efficiently unveil substructures in complex networks, giving birth to the field of community discovery. A novel and fascinating problem started capturing researcher interest recently: the identification of evolving communities. Dynamic networks can be used to model the evolution of a system: nodes and edges are mutable, and their presence, or absence, deeply impacts the community structure that composes them. This survey aims to present the distinctive features and challenges of dynamic community discovery and propose a classification of published approaches. As a “user manual,” this work organizes state-of-the-art methodologies into a taxonomy, based on their rationale, and their specific instantiation. Given a definition of network dynamics, desired community characteristics, and analytical needs, this survey will support researchers to identify the set of approaches that best fit their needs. The proposed classification could also help researchers choose in which direction to orient their future research.
Graph models provide an understanding of the dynamics of network formation and evolution; as a direct consequence, synthesizing graphs having controlled topology and planted partitions has been often identified as a strategy to describe benchmarks able to assess the performances of community discovery algorithm. However, one relevant aspect of real-world networks has been ignored by benchmarks proposed so far: community dynamics. As time goes by network communities rise, fall and may interact with each other generating merges and splits. Indeed, during the last decade dynamic community discovery has become a very active research field: in order to provide a coherent environment to test novel algorithms aimed at identifying mutable network partitions we introduce RDyn, an approach able to generates dynamic networks along with time-dependent ground-truth partitions having tunable quality.
Nowadays, online social networks represent privileged playgrounds that enable researchers to study, characterize and understand complex human behaviors. Social Network Analysis, commonly known as SNA, is the multidisciplinary field of research under which researchers of different backgrounds perform their studies: one of the hottest topics in such diversified context is indeed Community Discovery. Clustering individuals, whose relations are described by a networked structure, into homogeneous communities is a complex task required by several analytical processes. Moreover, due to the user-centric and dynamic nature of online social services, during the last decades, particular emphasis was dedicated to the definition of node-centric, overlapping and evolutive Community Discovery methodologies. In this paper we provide a comprehensive and concise review of the main results, both algorithmic and analytical, we obtained in this field. Moreover, to better underline the rationale behind our research activity on Community Discovery, in this work we provide a synthetic review of the relevant literature, discussing not only methodological results but also analytical ones.
Conference Paper
We propose ego-splitting, a new framework for detecting clusters in complex networks which leverage the local structures known as ego-nets (i.e. the subgraph induced by the neighborhood of each node) to de-couple overlapping clusters. Ego-splitting is a highly scalable and flexible framework, with provable theoretical guarantees, that reduces the complex overlapping clustering problem to a simpler and more amenable non-overlapping (partitioning) problem. We can scale community detection to graphs with tens of billions of edges and outperform previous solutions based on ego-nets analysis. More precisely, our framework works in two steps: a local ego-net analysis phase, and a global graph partitioning phase. In the local step, we first partition the nodes' ego-nets using a partitioning algorithm. We then use the computed clusters to split each node into its persona nodes that represent the instantiations of the node in its communities. Finally, in the global step, we partition the newly created graph to obtain an overlapping clustering of the original graph.