WLNI-LPA: Detecting Overlapping Communities in Attributed

Networks based on Label Propagation Process

Imen Ben El Kouni

1,2 a

, Wafa Karoui

1,3 b

and Lotﬁ Ben Romdhane

1,2 c

Universit

e de Sousse, Laboratoire MARS LR17ES05, ISITCom, 4011, Sousse, Tunisia

Universit

e de Sousse, ISITCom, 4011, Sousse, Tunisia

Universit

e de Tunis El Manar, Institut Superieur d’Informatique, 2080, Tunis, Tunisia

Keywords:

Attributed Networks, Overlapping Community Detection, Node Similarity, Weighted Graph.

Abstract:

Several networks are enriched by two types of information: the network topology and the attributes informa-

tion about each node. Such graphs are typically called attributed networks, where the attributes are always

as important as the topological structure. In these attributed networks, community detection is a critical task

that aims to discover groups of similar users. However, the majority of the existing community detection

methods in attributed networks were created to identify separated groups in attributed networks. Therefore,

detecting overlapping communities using a combination of nodes attributes and topological structure is chal-

lenging. In this paper, we propose an algorithm, called WLNI-LPA, based on label propagation for detecting

efﬁcient community structure in the attributed network. WLNI-LPA is an extension of LPA that combines

node importance, attributes information, and topology structure to improve the quality of graph partition. In

the experiments, we validate the performance of our method on synthetic weighted networks. Also, a part of

the experiment focuses on the impact of detecting signiﬁcantly overlapping communities in the recommender

system to improve the quality of recommendation.

1 INTRODUCTION

Network analysis has become a hot topic in recent

years due to the rapid growth of real-world networks

such as social networks (Leng and Jiang, 2016; Meo

et al., 2014). It can found in a wide variety of con-

texts, for example, model friendships and acquain-

tances in a social context; in biology, networks cap-

ture metabolic processes in the organism (Garza and

Schaeffer, 2019). A crucial task in network analy-

sis is group identiﬁcation, which is generally known

as a group of nodes with large internal connections

and minimal external connections. For example,

in protein-protein interaction networks, communities

refer to functional modules of interacting proteins.

Thus far, a large number of community-detection al-

gorithms have been proposed, and many of them

have successfully addressed the various aspects of the

community-detection issue. The attributed network is

a common community detection framework scenario,

https://orcid.org/0000-0003-3240-9647

https://orcid.org/0000-0002-5311-0655

https://orcid.org/0000-0003-2163-5809

in which nodes have attributes. In social networks,

for example, the users can be characterized by sev-

eral attributes such as gender, occupation, and hob-

bies. In attributed networks, community identiﬁcation

requires both network topology and nodes attributes

examination (Huang et al., 2016). Many commu-

nity detection studies, which only consider network

structure details, are not adaptable for attributed net-

works. The label propagation algorithm is one of the

most efﬁcient community detection approaches where

each node of the network is identiﬁed by a label.

Exploring nodes attributes during the label propaga-

tion process is technically challenging because label

propagation-based methods must assign one or more

labels to each node of the network. In general, com-

munities can be divided into two types: overlapping

and non-overlapping. Many existing community de-

tection algorithms in attributed network can solve the

non-overlapping community detection problem (Zhou

et al., 2009; Ruan et al., 2013). For that, studies start

to develop overlapping community algorithms in the

attributed networks to especially solve this problem

(Sun et al., 2012; Xu et al., 2012). But, the problem

is still open because these methods’ applicability is

408

El Kouni, I., Karoui, W. and Ben Romdhane, L.

WLNI-LPA: Detecting Overlapping Communities in Attributed Networks based on Label Propagation Process.

DOI: 10.5220/0010605904080416

In Proceedings of the 16th International Conference on Software Technologies (ICSOFT 2021), pages 408-416

ISBN: 978-989-758-523-4

restricted which means that are not able to handle all

types or size of networks.

In this paper, we propose an overlapping community

detection algorithm based on label propagation pro-

cess, called WLNI-LPA for attributed networks. This

algorithm tackles the issue of analyzing attributed

networks using label propagation. WLNI-LPA com-

bines the network’s topology information and node at-

tributes to ameliorate the accuracy of detection. Our

method is an extension of NI-LPA algorithm to de-

tect overlapping communities for attributed networks.

The original NI-LPA allows a node to contain more

than one label and can conveniently leverage these la-

bels in the label propagation step. In addition to the

propagation of labels, the idea of this work is to im-

prove the coefﬁcient of each label by attributes infor-

mation inherited from nodes. Moreover, WLNI-LPA

requires neither a pre-deﬁned objective function nor

prior information about the size of communities.

The following is a list of our major contributions:

• We propose an extension of the NI-LPA overlap-

ping community detection algorithm, which con-

siders both the topology of the network and at-

tribute information, to uncover communities in at-

tributed networks.

• We combine the node attributes information, the

node importance, and the multi-label propagation

by weighting the edges to add new knowledge

to identify overlapping communities in attributed

networks.

• The tests in synthetic networks show that our pro-

posed WLNI-LPA ameliorates the detection pro-

cess. Also, an application in the recommender

system area shows the efﬁciency of our approach

to detect overlapping communities and improve

the quality of recommendation.

2 RELATED WORK

In this section, we introduce the community detec-

tion issue and give a classiﬁcation of the different ap-

proaches. Then, we concentrate on the methods based

on the label propagation process.

2.1 Community Detection

One of the most important concepts in complex net-

work research is community detection. It helps to

analyze networks and understand behavior grouping.

Community detection is to group nodes into differ-

ent groups, where nodes in the same group are more

linked with each other than with nodes outside the

group. This area consists of creating cohesive groups

within networks, groups of friends in social networks,

or groups of web pages with the same subject in

knowledge networks. Other tasks are made possi-

ble by identifying such groups like marketing (Weng

et al., 2013), recommendation (Zheng and Wang,

2018; Zheng et al., 2019) and document organization

(Hachaj and Ogiela, 2017; Grineva et al., 2009). In

recent years, researchers have proposed a variety of

methods for identifying communities. Newman and

Girvan proposed a quality function called modular-

ity and optimized it for community detection (New-

man and Girvan, 2004). Since then, modularity-based

approaches that maximize the modularity value have

been commonly used to detect groups. However,

Lancichinetti and Fortunato (Lancichinetti and Fortu-

nato, 2011) demonstrated that the modularity measure

has a limited resolution and suffers extreme degra-

dation. Also, they propose a local ﬁtness maximiza-

tion approach for community detection. Lancichinetti

et al. (Lancichinetti et al., 2009) exploit the hierar-

chy of communities to apply a ﬁtness maximization

method for overlapping communities’ detection. In

random walk methods, the concept is based on the

fact that walks appear to be trapped in densely areas

of a network (Pons and Latapy, 2005). The most com-

mon random walk approach is Infomap, which is a

ﬂow-based group detection algorithm that combines

information-theoretic techniques and random walks

(Rosvall and Bergstrom, 2008). Label propagation

is an intensively studied issue in the ﬁeld of commu-

nity detection proposed by Raghavan et al. (Ragha-

van et al., 2007). This approach may be appropriate

for partitioning large networks in real time.

2.2 Label Propagation based

Community Detection

Label propagation algorithm (LPA) is a popular and

fast method for community detection. Initially, ev-

ery node is identiﬁed with unique labels. The next

step is for each node to update its label with the label

that is most frequently used by its neighbors. When

many neighbor labels are similarly frequent, it selects

at random from the most frequent labels. The process

of label propagation is repeated until all nodes with

the same label are grouped into a single community.

LPA is a simple, unsupervised near-linear algorithm

with no parameters that require no previous informa-

tion about the size and number of communities. Since

there is a random factor in LPA when there is more

than one most frequent label, different results can be

obtained after several runs. Various improvements to

LPA have been applied in recent years, in order to

WLNI-LPA: Detecting Overlapping Communities in Attributed Networks based on Label Propagation Process

409

enhance its stability and robustness. COPRA (Gre-

gory, 2010) algorithm with each vertex is assigned by

pairs (c, b), where c is a community identiﬁer and b

is a belonging coefﬁcient. BMLPA (Wu et al., 2012)

requires that one vertex’s community identiﬁers have

balanced belonging coefﬁcients, allowing nodes to

belong to any number of communities without impos-

ing a high number set by COPRA. SLPA (Xie et al.,

2011) employs a dynamic speaker-listener interaction

mechanism to maintain that each node can own sev-

eral labels. An extension of SLPA called WLPA (Hu,

2013) which improves it by adding a similarity be-

tween any two nodes focused on the labels they got

during label propagation. LPA-S (Li et al., 2017) is

proposed using label propagation and similarity to de-

velop a stepping community detection algorithm. LP-

LPA (Berahmand and Bouyer, 2018) algorithm im-

proves LPA by computing the link strength and node’s

label inﬂuence values and it processes label updating

according to the highest label inﬂuence value among

neighbors. AntLP (Hosseini and Rezvanian, 2020) is

an improved version of LPA that assigns weights for

edges based on several similarity indices, then, uses

ant colony optimization to propagate labels and opti-

mize modularity measure. NI-LPA (El Kouni et al.,

2020) is an extension that adopts LPA strategy to al-

low a node to contain a set of labels and simulates

a special propagation and ﬁltering process using in-

formation deduced from the structural properties of

nodes. On the other hand, these label propagation-

based methods are unable to address attributed net-

works because they all ignore nodes attributes infor-

mation during the propagation process. However, our

work uses the structural and semantic attribute infor-

mation of nodes, making it well adapted to attribute

networks.

3 PROPOSED METHOD

In this section, the problem deﬁnition has been de-

scribed. Then, we deﬁne the proposed weighted algo-

rithm called WLNI-LPA to detect overlapping com-

munities in weighted networks.

3.1 Problem Deﬁnition

Given a graph G = (V, E,L,W ) be an attributed net-

work. V = {v

,...,v

} is the set of n nodes,

E v V ∗ V is the set of edges, L = {l

,...,l

} is

the set of nodes attributes, and W is edge weight

between two nodes. Overlapping community detec-

tion is to partition the network G into k communities

C = {c

,...,c

} that satisfy the following criteria:

• Based on network structure, the nodes in the same

group are closely connected, while the nodes in

different communities are sparsely connected.

• Based on nodes attributes, nodes in the same

group have similar attribute values, while nodes

in different communities have different values.

• Based on overlap concept, node can allow to more

than one community c

3.2 Contribution

In our work, we propose to improve the NI-LPA

algorithm to be able to detect overlapping commu-

nities in attributed networks. On one hand, NI-LPA

focuses on both topology information and the role or

importance of the node in the network. On the other

hand, this algorithm maintains the simplicity of the

original LPA and obtains accurate results in large and

complex networks. For that, we show that NILPA

is an appropriate method to become adaptable for

attributed networks with large size, diverse attribute

nodes, and different typologies of the graph.

As for any algorithm of detection in attributed net-

works, the main topic is the use comprehensively of

structural information and nodes attributes together.

For example, in social networks, the proﬁles of users

can be regarded as nodes attributes. As illustrated in

Figure 1, each user is identiﬁed by his age, gender,

and occupation.

Figure 1: Social network with nodes attributes information.

Therefore, the proposed algorithm transforms the

nodes attributes information to a weight of edge to

compare the node in terms of semantic similarity. In

fact, this method needs a step of weighted network

construction. Then, the propagation phase will be im-

proved to take into consideration both the node impor-

tance and the link weighting. The proposed algorithm

is able to detect overlapping communities in attributed

networks.

ICSOFT 2021 - 16th International Conference on Software Technologies

410

3.3 Weighted Graph Construction

To integrate the information of nodes attributes, we

propose to transform the original unweighted net-

work to weighted networks where the weights of links

represent the similarity between nodes based on at-

tributes information. Therefore, we use the attributes

of each two node to calculate the similarity values be-

tween a pair of nodes. Then, affect this measure as a

weight of the link which connects these nodes. In this

work, we deﬁne the similarity between node u and

node v as the following:

Sim(u,v) =

∑

i=1

Sim(u,v)

(1)

where Nb indicates the number of attributes.

Figure 2 shows an explication of graph construc-

tion algorithm to transform attributed graph to a

weighted network. In this example, the attributes in-

formation is the tags put by the ﬁlm viewer. This ex-

ample demonstrates that the weighted network is eas-

ier to use in the propagation process.

Figure 2: Transformation from attributed graph to weighted

network.

3.4 Weighted NI-LPA

We propose an efﬁcient algorithm which is an ex-

tension of NI-LPA (El Kouni et al., 2020) to detect

overlapping communities in the attributed network.

In fact, the proposed algorithm, called WLNI-LPA,

is described in Algorithm 1. The step of weighted

network construction takes an unweighted attributed

network as input and creates a new weighted network

using the nodes attributes information to give link

weight as a similarity measure between two nodes.

For That, our algorithm takes as input the weighted

network and discovers a set of overlapping communi-

ties. It consists of three main components: initializa-

tion, propagation, and ﬁltering. Initially, we consider

each node as a community. The concept of overlap

means that each vertex may belong to more than one

community. For that, to ﬁnd overlapping communi-

ties, we must allow a node to contain many commu-

nity identiﬁers. In fact, a node is characterized by its

label, its topological importance, and its connections

with the others node. Similar to NI-LPA, we calculate

the importance of each node based on its degree and

coefﬁcient of clustering. Besides, the connection of

this node and its neighbors is assigned as a weight to

the link that connects with another node. The higher

the node importance and the link weight is, the higher

this label will be dominant in the propagation phase.

In the label propagation process of Lines 10–18, we

iterate the nodes labels Nb times. For each iteration i,

we update the labels of all nodes based on the labels

at previous iteration and the labels of its neighbors

updated in the current iteration. The node that is be-

ing processed receives a set of labels from its neigh-

bors and aggregates the scores of the labels associated

with the neighbors such as line 13 of the algorithm.

At the end of the propagation phase, each node con-

tains a set of labels with their scores. But, some of

these labels are poor in contrast to the other labels’

coefﬁcients. For that, lines 19-23 describe the simple

ﬁltering process. In this stage, we compare the co-

efﬁcient of each node with the threshold and remove

those who are less than it. The threshold is ﬁxed to

0.4 which means that any coefﬁcient less than 0.4 is

considered poor.

4 EXPERIMENTS

This section reports experimental settings and results.

On the one hand, we evaluate the proposed algo-

rithm in synthetic weighted networks with different

settings. On the other hand, we apply this algorithm

to evaluate its effectiveness to detect groups of similar

users in order to improve the quality of recommenda-

tion.

4.1 Experiments on Synthetic Networks

To evaluate the efﬁcacy of our proposed WLNI-LPA,

we carried out experiments on artiﬁcial networks us-

ing LFR benchmarks (Lancichinetti and Fortunato,

2009) to generate weighted networks with overlap-

ping communities. The LFR benchmark enables

the generation of weighted networks with power-law

node degree and community size distributions. For

the accuracy evaluation, we generate a group of the

dataset where N equal to 5000 and we varied the edge

weight and mixing parameter for topology. The pa-

rameter settings are shown in Table 1.

In addition, we compare the performance of our com-

munity detection algorithm (denoted as WLNI-LPA)

with three algorithms which can detect overlapping

communities for weighted networks.

• Conductance (Lu et al., 2014): characterizes

nodes in communities based on two metrics intra-

centrality and inter-centrality.

• COPRA (Gregory, 2010): detects communities

WLNI-LPA: Detecting Overlapping Communities in Attributed Networks based on Label Propagation Process

411

Algorithm 1: Weighted NI-LPA.

Data: A Weighted network G, number of

iterations Nb

Result: Overlapping communities

1 Map ← a new list with n empty dictionaries

foreach e in E do

2 u ← e.sourceNode

3 v ← e.targetNode

4 w ← e.weight

5 Initialize each node with a unique label

= x

6 Calculate importance of all the nodes

7 Sort nodes according to their degree in

descending order

8 for i=1 To Nb do

9 foreach label in List(u) do

10 if label is in List(v) then

11 List(u).label ←

List(v).label + w ∗

importance(v)

12 else

13 List(u).label ←

w ∗ importance(v)

14 end

15 end

16 end

17 end

18 Map ← List(u)

19 foreach Label with coefﬁcient b in Map do

20 if b < Threshold then

21 delete label

22 end

23 end

24 return Map

on weighted networks based on label propagation

process.

• Strength (Chen et al., 2010): exploits belonging

degree and node strength to detect the overlapping

community structure.

Since it is hard to verify the detected communities

for real networks, the evaluation of these methods is

based mainly on synthetic networks.

As metric, we use the normalized mutual information

NMI (Lancichinetti et al., 2009) deﬁned by Equation

2 to compare communities with ground-truth parti-

tion. The higher NMI values mean good partition.

norm

(X : Y ) =

H(X) + H(Y ) − H(X,Y )

(H(X) + H(Y ))/2

(2)

where H(X), (H(Y)), is the entropy of the random

variable X, (Y), assigned to the partition C0, (C00).

Threshold Deﬁnition. In the ﬁltering step, the co-

efﬁcient of each node is compared with the threshold

and remove those who are less than it. The choice

of this value is clearly explained by the fact that any

coefﬁcient less than this threshold is considered poor.

As a consequence, in this section, we will attempt to

test distinct values experimentally. The parameters

of networks are as follows: mut = 0.2; muw = 0.2;

k = 10; O

= 2; O

= 10%. Table 2 reports the re-

sults. We can conclude that our algorithm gives the

best value of NMI when the threshold is 0.4, and the

quality of detection decreases when we consider an-

other value.

Evaluation of WLNI-LPA with Different Settings.

Figure 3 presents the NMI values given by our algo-

rithm with the different average degrees on the LFR

networks. We test WLNI-LPA with three categories:

when the mixing parameter mut equal to 0.2; 0.3; and

0.5. As the value of mut increases, the network be-

comes much more complex, and the boundaries be-

tween communities become more unclear. We show

in this ﬁgure that all the values are between 0.96 and

1 which means that WLNI-LPA successfully provides

a partition very close to the exact partition. This is

mainly because our algorithm collects more and more

useful information (degree measure, clustering coefﬁ-

cient, and node attributes) about the node to propagate

labels.

Figure 3: The NMI measures as a function of the average

degree in a weighted network with 5000 nodes.

Figure 4: The NMI measures as a function of the percentage

of the overlapping node in a weighted network with 5000

nodes.

As illustrated in Figure 4, we test our method with

two constraints which are the percentage of overlap-

ping nodes from 10% to 40% and the variation of

ICSOFT 2021 - 16th International Conference on Software Technologies

412

Table 1: Parameter settings of benchmarks.

Param value Description

N 5000 Number of nodes

k 15 - 25 Average node degree

maxk 50 Max node degree

mut 0.2 - 0.5 Mixing parameter for topology

muw 0.2 - 0.5 Mixing parameter for edge weight

minc 20 Minimum for community sizes

maxc 50 Maximum for community sizes

On 10-40 Number of overlapping nodes

Om 2 Number of communities of overlapping nodes

Table 2: NMI results for test with different threshold.

Network 0.2 0,3 0,4 0,5 0,6

5000 0,81 0,9 1 0,96 0,72

Figure 5: The quality in terms of NMI of the community structure detected by Conductance (Lu et al., 2014), COPRA

(Gregory, 2010), Strength (Chen et al., 2010) and WLNI-LPA for different parameter settings.

Test x: N=5000, µ

= 0.1, µ

= 0.1

Test y: N=5000, µ

= 0.3, µ

= 0.3.

the mixing parameter for edge weight. The results

demonstrate that WLNI-LPA detect overlapping com-

munities in all the cases. This indicates that our algo-

rithm can propagates labels by considering topologi-

cal and attributes information and then, ﬁlter useless

labels to ﬁnd more accurate communities with differ-

ent numbers of overlapping nodes.

Figure 5 shows the results of our algorithm com-

pared to three other algorithms from the literature

for different parameter settings. We can see the per-

formance all these approaches degrade dramatically

when the overlapping fraction increase from 10% to

60%. The NMI values given by WLNI-LPA indi-

cate that it is able to detect good partition in artiﬁ-

cial weighted networks compared to other methods.

WLNI-LPA has a very stable performance compared

to strength and COPRA even with the increase of

overlapping nodes fraction. However, WLNI-LPA

and conductance give closer partitions to exact ones.

4.2 Application to Social Recommender

System

Recently, the socialized recommendation has become

one of the most common methods of recommenda-

tion in a variety of recommender systems used in

areas such as e-commerce, social media sites, and

web search engines (Ikeda et al., 2013). Integrat-

ing detection community algorithm in social recom-

mender system becomes an important challenge stud-

ied by many authors (Gasparetti et al., 2020; Bous-

saadi et al., 2020; Ai et al., 2019). In social recom-

mender systems, it’s crucial to ﬁnd users that have

similar interests in order to provide recommendations.

To address these problems, user proﬁle techniques are

used to reﬂect users’ interests and detect similar users.

Therefore, integrating the community detection pro-

cess in the recommender system can enhance recom-

mendations. In this paper, since WLNI-LPA is a com-

munity detection algorithm speciﬁc to attributed net-

works, we use it as a step in the recommender sys-

tem algorithm to detect similar users. To evaluate

WLNI-LPA: Detecting Overlapping Communities in Attributed Networks based on Label Propagation Process

413

Table 3: Datasets used in recommender system.

Datasets Users Items R-scale demographic information

Flickr 1000 500 1-5 age, location, gender

Book crossing 2000 1000 1-10 age, location

Figure 6: Performance comparison (Precision, Recall, and F-measure) on attributed and non-attributed networks using Flickr

datasets.

Figure 7: Performance comparison (Precision, Recall, and F-measure) on attributed and non-attributed networks using Book

crossing datasets.

Figure 8: Performance comparison (Diversity, rate Coverage) on attributed and non-attributed networks using Book crossing

datasets.

the performance of this method using WLNI-LPA, we

use two known datasets in the recommender system

ﬁeld and described in Table 3. Flickr is an image-

sharing network in which nodes represent users and

links represent relationships between users. Book-

crossing contains a set of users (with demographic in-

formation) providing ratings about a set of books.

We used precision, recall, F1 measure, diversity,

and rate coverage (RC) to evaluate the performance

of recommendation.

Let TP is the recommendations generated by the algo-

rithm that users like deﬁned as True Positive, and the

others those not like by the users are deﬁned as False

Positive (FP). The items that are not recommended

and the uses not like are deﬁned as True Negative

(TN), and those not recommended by the algorithm

but liked by users are deﬁned as False Negative (FN).

These metrics are calculated as follows:

Prec =

∑

i=1

(

T P

T P + FP

) (3)

Rec =

∑

i=1

(

T P

T P + FN

) (4)

F − measure = 2 ∗

Prec ∗ Rec

Prec + Rec

(5)

Diversity =|

[

i=1,..,n

(i) | (6)

where L

(i) is the Top-N items in the recommenda-

tion list for the i

user, and n is the number of users.

RC =

number o f predicted ratings

number o f all ratings

(7)

ICSOFT 2021 - 16th International Conference on Software Technologies

414

The objective of this recommender system is to

recommend items to users based on the similarity be-

tween them. For that, we apply our algorithm WLNI-

LPA to detect a group of similar users with overlap.

Thus, improving the algorithm of detection means en-

hancing the quality of recommendation.

In fact, the demographic information of users will

be considered as nodes attributes. To demonstrate

how the nodes attributes affect the quality of recom-

mendations, we use our proposed algorithm WLNI-

LPA in the social recommendation process with and

without attributes also to see how well it utilizes the

attribute information in the label propagation.

Figure 6 and 7 recapitulate the experimental re-

sults in both datasets. The results show that run-

ning WLNI-LPA on an attributed network achieves

great promotion in the community detection task in all

cases (Top 5 to 20) and even in two datasets with dif-

ferent densities. The scores given when using WLNI-

LPA in the attributed network is always better than in

non attributed network in terms of quality of results

which means that the attributed information improves

the quality of detecting similar users and can bring

satisfaction about recommended items. In the other

part, the test on Book-crossing datasets with consid-

ering attributed network achieves maximum scores of

diversity and RC than non attributed network. The

good partition of the network in communities affects

the result of predicted rating produced by a recom-

mender system.

5 CONCLUSION

In this paper, we examined the community detection

analyses in attributed graphs. We propose WLNI-

LPA, an accurate method for detecting overlapping

communities in attributed networks as only a few

works are considering the overlap concept. Our

method combines the nodes attributes with the net-

work topology based on the label propagation pro-

cess to improve the quality of graph partition. WLNI-

LPA transforms the attributed network to a weighted

graph where the attributes information will be repre-

sented as link weight between nodes. The experimen-

tal tests are divided into two types: ﬁrstly, in an ar-

tiﬁcial weighted network and secondly by integrating

WLNI-LPA in a recommender system to detect sim-

ilar users. The results show that our algorithm can

effectively discover the overlapping communities on

one hand and improve the quality of recommendation

on the other hand.

Future work can be conducted in various directions.

First, to improve the accuracy by optimizing the

weights of node attributes. Second, to propose tempo-

ral similarity measurements which consider the time

factor to detect communities in dynamic attributed

networks. Third, since real-life problems models

use dynamic real-world networks graphs to repre-

sent entities and the relations between them, we must

study these dynamic networks evolution. As a solu-

tion, on one hand, we can add some information into

group detection problems that are not intrinsic to the

graph structure; on the other hand, we can consider

the ability of the members to communicate among

them. Indeed, the community would be more ho-

mogeneous if we take into consideration different ex-

changes between members like public or private mes-

sages, videos, photos, hypertext links, or games.

REFERENCES

Ai, J., Liu, Y., Su, Z., Zhang, H., and Zhao, F. (2019). Link

prediction in recommender systems based on multi-

factor network modeling and community detection.

EPL (Europhysics Letters), 126(3):38003.

Berahmand, K. and Bouyer, A. (2018). Lp-lpa: A link

inﬂuence-based label propagation algorithm for dis-

covering community structures in networks. Interna-

tional Journal of Modern Physics B, 32(06):1850062.

Boussaadi, S., Aliane, H., Abdeldjalil, O., Houari, D.,

and Djoumagh, M. (2020). Recommender systems

based on detection community in academic social

network. In 2020 International Multi-Conference

on:“Organization of Knowledge and Advanced Tech-

nologies”(OCTA), pages 1–7. IEEE.

Chen, D., Shang, M., Lv, Z., and Fu, Y. (2010). Detecting

overlapping communities of weighted networks via a

local algorithm. Physica A: Statistical Mechanics and

its Applications, 389(19):4177–4187.

El Kouni, I. B., Karoui, W., and Romdhane, L. B. (2020).

Node importance based label propagation algorithm

for overlapping community detection in networks. Ex-

pert Systems with Applications, 162:113020.

Garza, S. E. and Schaeffer, S. E. (2019). Community detec-

tion with the label propagation algorithm: A survey.

Physica A: Statistical Mechanics and its Applications,

534:122058.

Gasparetti, F., Sansonetti, G., and Micarelli, A. (2020).

Community detection in social recommender systems:

a survey. Applied Intelligence, pages 1–21.

Gregory, S. (2010). Finding overlapping communities

in networks by label propagation. New journal of

Physics, 12(10):103018.

Grineva, M., Grinev, M., and Lizorkin, D. (2009). Extract-

ing key terms from noisy and multitheme documents.

In Proceedings of the 18th international conference on

World wide web, pages 661–670.

Hachaj, T. and Ogiela, M. R. (2017). Clustering of trend-

ing topics in microblogging posts: A graphbased

WLNI-LPA: Detecting Overlapping Communities in Attributed Networks based on Label Propagation Process

415

approach. Future Generation Computer Systems,

67:297–304.

Hosseini, R. and Rezvanian, A. (2020). Antlp: ant-based

label propagation algorithm for community detection

in social networks. CAAI Transactions on Intelligence

Technology, 5(1):34–41.

Hu, W. (2013). Finding statistically signiﬁcant communities

in networks with weighted label propagation.

Huang, X., Cheng, H., and Yu, J. X. (2016). Attributed com-

munity analysis: Global and ego-centric views. IEEE

Data Eng. Bull., 39(3):29–40.

Ikeda, K., Hattori, G., Ono, C., Asoh, H., and Higashino,

T. (2013). Twitter user proﬁling based on text and

community mining for market analysis. Knowledge-

Based Systems, 51:35–47.

Lancichinetti, A. and Fortunato, S. (2009). Benchmarks for

testing community detection algorithms on directed

and weighted graphs with overlapping communities.

Physical Review E, 80(1):016118.

Lancichinetti, A. and Fortunato, S. (2011). Limits of modu-

larity maximization in community detection. Physical

review E, 84(6):066122.

Lancichinetti, A., Fortunato, S., and Kert

esz, J. (2009).

Detecting the overlapping and hierarchical commu-

nity structure in complex networks. New journal of

physics, 11(3):033015.

Leng, J. and Jiang, P. (2016). Mining and matching rela-

tionships from interaction contexts in a social manu-

facturing paradigm. IEEE Transactions on Systems,

Man, and Cybernetics: Systems, 47(2):276–288.

Li, W., Huang, C., Wang, M., and Chen, X. (2017). Step-

ping community detection algorithm based on label

propagation and similarity. Physica A: Statistical Me-

chanics and its Applications, 472:145–155.

Lu, Z., Sun, X., Wen, Y., Cao, G., and La Porta, T. (2014).

Algorithms and applications for community detection

in weighted networks. IEEE Transactions on Parallel

and Distributed Systems, 26(11):2916–2926.

Meo, P. d., Ferrara, E., Abel, F., Aroyo, L., and Houben,

G.-J. (2014). Analyzing user behavior across social

sharing environments. ACM Transactions on Intelli-

gent Systems and Technology (TIST), 5(1):1–31.

Newman, M. E. and Girvan, M. (2004). Finding and eval-

uating community structure in networks. Physical re-

view E, 69(2):026113.

Pons, P. and Latapy, M. (2005). Computing communities

in large networks using random walks. In Interna-

tional symposium on computer and information sci-

ences, pages 284–293. Springer.

Raghavan, U. N., Albert, R., and Kumara, S. (2007).

Near linear time algorithm to detect community struc-

tures in large-scale networks. Physical review E,

76(3):036106.

Rosvall, M. and Bergstrom, C. T. (2008). Maps of random

walks on complex networks reveal community struc-

ture. Proceedings of the National Academy of Sci-

ences, 105(4):1118–1123.

Ruan, Y., Fuhry, D., and Parthasarathy, S. (2013). Efﬁcient

community detection in large networks using content

and links. In Proceedings of the 22nd international

conference on World Wide Web, pages 1089–1098.

Sun, Y., Aggarwal, C. C., and Han, J. (2012). Rela-

tion strength-aware clustering of heterogeneous infor-

mation networks with incomplete attributes. arXiv

preprint arXiv:1201.6563.

Weng, L., Menczer, F., and Ahn, Y.-Y. (2013). Virality pre-

diction and community structure in social networks.

Scientiﬁc reports, 3(1):1–6.

Wu, Z.-H., Lin, Y.-F., Gregory, S., Wan, H.-Y., and

Tian, S.-F. (2012). Balanced multi-label propagation

for overlapping community detection in social net-

works. Journal of Computer Science and Technology,

27(3):468–479.

Xie, J., Szymanski, B. K., and Liu, X. (2011). Slpa: Un-

covering overlapping communities in social networks

via a speaker-listener interaction dynamic process. In

2011 ieee 11th international conference on data min-

ing workshops, pages 344–349. IEEE.

Xu, Z., Ke, Y., Wang, Y., Cheng, H., and Cheng, J. (2012).

A model-based approach to attributed graph cluster-

ing. In Proceedings of the 2012 ACM SIGMOD in-

ternational conference on management of data, pages

505–516.

Zheng, J., Wang, S., Li, D., and Zhang, B. (2019). Per-

sonalized recommendation based on hierarchical in-

terest overlapping community. Information Sciences,

479:55–75.

Zheng, J. and Wang, Y. (2018). Personalized recommenda-

tions based on sentimental interest community detec-

tion. Scientiﬁc Programming, 2018.

Zhou, Y., Cheng, H., and Yu, J. X. (2009). Graph clustering

based on structural/attribute similarities. Proceedings

of the VLDB Endowment, 2(1):718–729.

ICSOFT 2021 - 16th International Conference on Software Technologies

416