1776 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 9, NO.

3, MAY/JUNE 2022

A Discrete Moth-Flame Optimization

With an l2-Norm Constraint for Network Clustering
Xianghua Li , Xin Qi, Xingjian Liu, Chao Gao , Member, IEEE, Zhen Wang , Fan Zhang, and
Jiming Liu , Fellow, IEEE

Abstract—Complex network clustering problems have been and mathematics, complex networks are often represented as a
gained great popularity and widespread researches recently, and graph where vertices and edges denote the objects and the rela-
plentiful optimization algorithms are aimed at this problem. Among
these methods, the optimization methods aiming at multiple tions between them in complex systems, respectively. The clus-
objectives can break the limitations (e.g., instability) of those ter structures are ubiquitous in networks, such as social clusters
optimizing single objective. However, one shortcoming stands out based on different backgrounds or interests in social net-
that these methods cannot balance the exploration and exploitation works [5], and subway stations cluster based on different passen-
well. In another sentence, it fails to optimize solutions on the basis of ger flows in subway networks [6]. Here, a cluster (or
the good solutions obtained so far. Inspired by nature, a new
optimized method, named multi-objective discrete moth-flame community) is defined as a collection of closely-connected sub-
optimization (DMFO) method is proposed to achieve such a set vertices, with sparse connections between different clus-
tradeoff. Specifically, we redefine the simple flame generation (SFG) ters [7]. The aim of network clustering is to detect cluster
and the spiral flight search (SFS) processes with network topology structures in a network, namely, to divide a network into many
structure to balance exploration and exploitation. Moreover, we subsets of vertices according to the properties of vertices.
present the DMFO in detail utilizing a Tchebycheff decomposition
method with an l2 -norm constraint on the direction vector (2-Tch). Network clustering can be essentially modeled as an optimi-
Besides that, experiments are taken on both synthetic and real- zation problem [8]. Generally, such a clustering optimization
world networks and the results demonstrate the high efficiency and problem is deemed to be NP-hard, and several optimization
promises of our DMFO when tackling dividing complex networks. algorithms have been proposed for solving this problem
Index Terms—Network clustering, multi-objective optimiza- recently [9]. Traditional single-objective methods tend to select
tion, discrete moth-flame optimization, decomposition. an optimal objective at first (such as the modularity Q [10], [11],
the modularity density [12] or the community score [13]) and
optimize such an objective by leveraging different strategies. A
I. INTRODUCTION quintessential example is that GA-Net optimizes the community
score by means of a genetic algorithm for network clustering in
C OMPLEX networks are omnipresent in reality, such as
biological networks [1], social networks [2], transportation
networks [3] and biochemical networks [4]. In computer science
social networks [13]. DCRO optimizes the modularity in a way
the chemical reaction process takes place for the acquisition of a
better clustering result [14]. Nevertheless, these methods share a
common shortcoming that they only take structure into account
Manuscript received March 7, 2021; revised September 2, 2021; accepted
February 17, 2022. Date of publication February 23, 2022; date of current ver- when formulating objective functions, even when the multiple
sion May 23, 2022. This work was supported in part by the National Science underlying structural attributes ask for solutions. Therefore, in
Fund for Distinguished Young Scholars under Grant 62025602, in part by the order to overcome such a shortcoming, the multi-objective meth-
National Natural Science Foundation of China under Grants 61976181,
U1803263, and 11931015, in part by Fok Ying-Tong Education Foundation under ods mining multiple characteristics of a network simultaneously
Grant 171105, in part by the Fundamental Research Funds for the Central Univer- have been applied to make network clustering more accurate and
sities under Grants D5000210827 and D5000210738, in part by the Natural Sci-
ence Foundation of Shaanxi Province under Grant 2022JM-325, and in part by
efficient [15]–[20]. For example, a multi-objective network clus-
Tencent Foundation and XPLORER PRIZE. Recommended for acceptance tering method (MOGA-Net) leverages the non-dominated sorting
by Dr. Gaoxi Xiao. (Corresponding authors: Zhen Wang; Chao Gao) genetic algorithm-II (NSGA-II) to improve both the community
Xianghua Li and Chao Gao are with the School of Artificial Intelligence,
Optics, and Electronics (iOPEN), Northwestern Polytechnical University,
score and community fitness simultaneously [20]. Although a
Xian 710072, China, and also with the College of Computer and Information discrete multi-objective particle swarm optimization (MODPSO)
Science & College of Software, Southwest University, Chongqing 400715, China method considers the kernel K-means (KKM) and Ratio
(e-mail: [email protected]; [email protected]).
Xin Qi, Xingjian Liu, and Fan Zhang are with the College of Computer and Cut (RC) to further improve the accuracy of the division
Information Science & College of Software, Southwest University, Chongqing results [15], [21], there is no clear accurate timing for transition
400715, China (e-mail: [email protected]; 2638666276@foxmail. between exploration and development. Therefore it is difficult
com; [email protected]).
Zhen Wang is with the School of Artificial Intelligence, Optics, and Elec- for multi-objective optimization algorithms to achieve a tradeoff
tronics (iOPEN), Northwestern Polytechnical University, Xian 710072, China between exploration and development in the calculation process.
(e-mail: [email protected]). In recent years, more and more meta-heuristic algorithms have
Jiming Liu is with the Department of Computer Science, Hong Kong Bap-
tist University, Hong Kong (e-mail: [email protected]). been applied to solve multi-objective optimization problems. For
Digital Object Identifier 10.1109/TNSE.2022.3153095 example, the monarch butterfly optimization (MBO) is a very
2327-4697 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tps://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DISCRETE MOTH-FLAME OPTIMIZATION WITH AN L_2-NORM CONSTRAINT FOR NETWORK CLUSTERING 1777

competitive metaheuristic algorithm, which is very easy to imple- optimal value. More specifically, based on the label-based repre-
ment because it only needs to fine-tune migration and adjust oper- sentation of individuals, we adjust the SFG and SFS processes to
ators [22]. Inspired by the cooperative behaviors and chasing extend a discrete MFO variant. On the basis of two optimization
styles of Harris hawks, a novel population based optimization functions which contradicts each other (i.e., KKM and RC), the
algorithm, called Harris hawks optimization (HHO), is proposed multi-objective optimization network clustering framework gets
to tackle different optimization tasks [23]. A new stochastic opti- optimized. The contributions of the paper are as follows:
mizer, which is called slime mould algorithm (SMA), is designed
For the balance between exploration and exploitation
based on the oscillation mode of slime mould [24]. The SMA when clustering networks, the discrete MFO variant
algorithm has several new features with a unique mathematical (denoted as DMFO) is extended for the first time. Such
model to form the optimal path for connecting foods with the a variant redefines the SFG and SFS processes based on
excellent exploratory ability and exploitation propensity. More- the features of networks.
over, A Hunger Games Search (HGS) is designed according to
To divide and conquer the network clustering problems,
the hunger-driven activities and behavioral choice of ani- the improved Tchebycheff method (2-Tch) is leveraged
mals [25]. The HGS algorithm incorporates the concept of hunger in our DMFO, which is first used to address network
into the feature process, in other words, an adaptive weight based clustering.
on the concept of hunger is designed and employed to simulate
Extensive comparative experiments on both synthetic
the effect of hunger on each search step. Furthermore, a general- and real-world networks validate the good performance
purpose metaheuristic method named moth search algorithm of our proposed DMFO.
(MSA) is proposed to generalize the phototaxis and Levy flights The rest of this paper is designed as follows. The background is
of moths [26]. The features mentioned above correspond to the given in Section II. Our approach for network clustering is dem-
exploitative and explorative process of an optimization algo- onstrated in Section III. Then, extensive experiments are imple-
rithm. The above-mentioned multi-objective optimization algo- mented to prove the superiority of the proposed DMFO algorithm
rithms have proved their superiority in detecting communities in in Section IV. Finally, conclusions are made in Section V.
complex networks. However, due to the disaster of dimensional-
ity, when these algorithms deal with complex optimization II. RELATED WORKS
problems with a large number of local optimal solutions, the con-
A. Problem Formulation
vergence speed is slow and there are easy to fall into the local
optimal. How to balance the exploration ability and the exploita- An undirected graph G ¼ ðV; EÞ is usually utilized to repre-
tion ability to obtain an ideal optimization result is still a big prob- sent a network, where V and E stand for a set of vertices and
lem that needs to be solved urgently. edges, correspondingly. A network cluster is defined as a
The moth-flame optimization (MFO) algorithm is a new sto- group of vertices, where the number of edges inside the same
chastic optimization algorithm [27]. With the simple flame gen- cluster is higher than that of edges between different clus-
eration (SFG) and the spiral flight search (SFS) processes [28] ters [7]. The purpose of network clustering is to divide the net-
merged into it, the MFO can get a good balance between explo- work into closely connected subgraphs Gi ði ¼ 1; . . . ; kÞ in G,
ration and exploitation. Considering these features, Xu et al. where G obeys:
[29] firstly propose a series of mutation-based MFO, such as
LGCMFO and GMFO. These MFO variants use mutation meth- [1ik Gi ¼ G and \1ik Gi ¼ ? (1)
ods such as Gaussian mutation (GM), Cauchy mutation (CM),
Levy mutation (LM) to balance the exploration and exploitation
In reality, network clustering needs to comprehensively
processes in order to improve the performance of the basic
consider the attributes between intra-community and inter-
MFO. Moreover, a double adaptive weight mechanism is intro-
community, which means that the detection process should
duced into the MFO algorithm [30], termed as WEMFO, which
consider the multiple optimization metrics. Currently, the
adjusts the search strategy adaptively in different periods of
widely used optimization metrics are RC and KKM [15],
algorithm, to make it more flexible between global search
which are a pair of conflicting objective functions, corre-
(diversification) and local search (intensification). Furthermore,
sponding to the sparse connection of inter-communities and
the CLSGMFO algorithm introduces two mechanisms, Gaussian
the close connection of intra-communities, respectively. In
mutation and chaotic local search, into the original MFO algo-
order to get an excellent clustering effect reflected by such a
rithm [31]. More specifically, Gaussian mutation operation
pair of metrics, the multi-objective optimization process is
increases the diversity of the moth population during the explo-
defined in (2). More specifically, there are N vertices and jEj
ration phase and the chaotic local search mechanism provides
edges in G, the adjacency matrix of G is A ¼ ðAij ÞNN . If
the original MFO with the ability to jump out of local optima.
there exists an edge between vertices vi and vj in a network,
However, these methods don’t address the impact of premature
Aij ¼1, otherwise, 0.
convergence on the network clustering.
In this paper, we propose a novel multiobjective discrete moth- 8 P
< f1 ¼ KKM ¼ 2ðN  cÞ  ci¼1 LðVi ;Vi Þ
fame optimization algorithm (DMFO) for processing network jVi j
min
: f2 ¼ RC ¼ P c L ð V ;V Þ (2)
clustering. By continuously adjusting the positions of the flames i i
i¼1 jVi j
and moths, our method can accurately approximate the global

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
1778 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 9, NO. 3, MAY/JUNE 2022

where c represents
P the number of cluster in a network. the convergence speed and computational performance of the
LðVi ; Vi Þ ¼ i2Vi ;j2Vi Aij means the internal link densities of network clustering algorithm.
nodes both coming from the same community i. In contrast, Moreover, some approaches, such as MOEA/D-Net [37],
P
LðVi ; V i Þ ¼ i2Vi ;j2V i Aij represents the external link densities APMOEA [38] and MODBSA/D [39], are proposed based on
of vertices. Specifically, Vi denote a set of nodes coming from the decomposition to explore solution spaces by minimizing two
the ith community while V i stands for a set of nodes which contradictory optimization objectives, i.e., the ratio cut (RC) and
belongs to any other communities except the ith community. negative ration association (NRA). In addition, two contradic-
tory optimization functions, i.e., Kernel K-Means (KKM) and
Ratio Cut (RC), are proposed to develop the discrete multi-objec-
B. Nature-Inspired Computing
tive particle swarm optimization (MODPSO) [15] algorithm for
Network clustering is a key field of complex network dividing networks. This set of objective functions has gained
research, which have been widely studied in recent years. great popularity as soon as proposed. For instance, a quantum-
Numerous nature-inspired optimization algorithms are aimed mechanism-based PSO algorithm with non-dominated sorting
at solving this problem [32]–[34]. These solutions roughly selection mechanism [16], a multi-objective particle swarm opti-
split into two types, i.e., single-objective optimization and mization algorithm (MOPSO-Net) [17], and a network reduc-
multi-objective optimization approaches. tion-based multi-objective evolutionary algorithm (RMOEA) [9],
Concerning the first category of single objective optimization, which all use these two objective functions to achieve the better
only one objective function is selected to be optimized. For solution of network clustering.
instance, Pizzuti developed a genetic algorithm (GA-Net) for However, the above-mentioned optimization algorithms,
network clustering in social networks [13] by optimizing a sim- whether single-objective or multi-objective, are easy to fall
ple and effective fitness function. Chang et al. developed a into the local optimum and it is difficult to achieve the balance
DCRO [14] method, which simulates the chemical reaction of exploration and development in the process of searching for
process, thereby optimizing the modularity to achieve network solutions, which leads to the loss of optimal solutions. The
clustering. A discrete fireworks algorithm (DMFWA) [12] is Moth Flame Optimization (MFO) algorithm is an intelligent
proposed to solve network clustering problems by maximizing optimization algorithm based on population. Its particles move
the modularity density function. in a spiral way around a better solution, which not only ensures
As for the second category about the multi-objective opti- that algorithm is unlikely to fall into the local optimum, but
mization algorithms, they optimize multiple objective func- also enhances the parallel optimization ability of algorithm and
tions simultaneously for network clustering, which can the balance ability of exploration in order to obtain better clus-
compensate for the shortcomings of single-objective optimi- tering solution. Currently, the MFO algorithm has achieved
zation algorithms. In the last few years, many efforts have excellent results in many practical engineering optimization
been devoted to the application of evolutionary algorithms problems [40]–[42].
(EAs) to the development of multi-objective optimization
algorithms. For example, based on the non-dominated sorting C. The Moth-Flame Optimization
genetic algorithm-II (NSGA-II) [20], a variety of multi-
objective network clustering methods have been developed. 1) Simple Flame Generation Process: Inspired by the nat-
Among them, the MOGA-Net algorithm maximizes intra- ural moth navigation mechanism, Mirjalili proposed a novel
connections and minimizes inter-connections simultaneously population-based stochastic optimization algorithm called the
based on the community score and community fitness (e.g., Moth-Flame Optimization (MFO) algorithm [27]. The MFO
two contradictory objective functions) [20], respectively. algorithm contains two populations, namely the moth popula-
Unlike only considering the structure of a network, MOGA- tion and the flame population. The entire space where the moth
@Net algorithm integrates the structure and composition flies is the solution space of the problem. Specifically, each
dimensions contained in the attribute network, and can obtain moth corresponds to a light source, which avoids the algorithm
a division result that is both well-connected and similar falling into the local optimum. When there are enough moths
nodes [35]. Moreover, Li et al. proposes a multi-agent and flames, the flight of moths can search most of the solution
genetic algorithm for community detection by designing a set space, thus ensuring the exploration ability of algorithm. In the
of effective neighborhood-based operators [36]. It uses the process of optimization, the number of flames decreases as the
matrix coding and node similarity to generate the initial pop- round of iterations increases, so that the moth can fully search
ulation, which enhances the diversity of initial individuals the neighborhood space of better solutions, and the utilization
and improves the efficiency of finding the optimal solution. ability of the algorithm is ensured.
In summary, these algorithms achieve superior results in net- At the same time, the MFO algorithm contains two key pro-
work clustering problems, however, due to the curse of cesses: Simple Flame Generation (SFG) and Spiral Flight
dimensionality, they converge slowly when dealing with Search (SFS). Among them, the purpose of SFG process is to
complex optimization problems with a large number of local generate the next generation of flames from the best set of moth
optimal solutions. Meanwhile, they do not address the prob- and flame populations obtained so far. Then, the SFS process
lem of balancing exploration and development in the process imitates the horizontal positioning behavior of moth, guiding
of searching for solutions, which is significantly in improving the moth to fly to the flame in a spiral shape, thereby updating

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DISCRETE MOTH-FLAME OPTIMIZATION WITH AN L_2-NORM CONSTRAINT FOR NETWORK CLUSTERING 1779

its position in the iterative process. Finally, by continuously

repeating SFG and SFS processes, MFO can find the optimal
solutions in the search space. For a D-dimension optimization
with n flames and moths, each flame and each moth can be
defined as Fi ¼ ½fi1 ; . . . ; fiD  and Mi ¼ ½mi1 ; . . . ; miD ; i 2
½1; n, respectively. The detailed whole process is as below.
First, when the number of iterations k is 1, fM1 ; . . . ; Mn g is
sorted into fS1 ; . . . ; Sn g according to their corresponding fit-
ness value in increasing order. Specifically, Si ; i 2 ½1; n cor- Fig. 1. An illustration of the label-based representation. (a) The true partition
responds to the ith individual in M after being sorted in and encode of a network topology, which can be divided into four clusters. (b)
increasing order, i.e., Fi is equal to Si , i 2 ½1; n. Then, when An individual decoding of a network. Nodes of the same color represent the
same division.
the iterative number k > 1, the Mi ðk  1Þ and Fi ðk  1Þ; i 2
½1; n is sorted increasingly by their corresponding fitness
value. Fi ðkÞ; i 2 ½1; n can be acquired by obtaining the pre- Section III-A introduces the representation of the discrete posi-
ceding n individuals in the merged Mi ðk  1Þ and Fi ðk  tion and the initialization of populations. Then, Section III-B
1Þ; i 2 ½1; n which are arranged in increasing order with their and Section III-C are used respectively to redesign the two pro-
corresponding fitness value. Here, Mi ðk  1Þ and Fi ðk  cesses of DMFO (i.e., SFG and SFS) to achieve a good balance
1Þ; i 2 ½1; n represent the positions of moths and flames in the between exploration and exploitation in network clustering.
ðk  1Þth iteration, respectively. Indeed, both Mi ðk  1Þ and Moreover, we present the DMFO framework based on two
Fi ðk  1Þ; i 2 ½1; n are solutions in the solution space. The newly designed processes using an improved Tchebycheff
difference between them is that Mi ðk  1Þ; i 2 ½1; n is the decomposition method in Section III-D. Finally, Section III-E
real search locations of moths, and Fi ðk  1Þ; i 2 ½1; n is the analyzes the computational complexity of DMFO method.
optimal positions in the moths and flames obtained up to now.
2) Spiral Flight Search Process: The spiral flight search A. Representation and Initialization of Individuals
(SFS) process is used to iteratively update moth positions
How to represent the discrete position is essentially the
when each moth flies spirally towards its corresponding flame.
problem of the encoding scheme for solutions, which is criti-
In the original MFO [27], the logarithmic spiral function is
cal to algorithms. This paper applies a label-based representa-
chosen to simulate the moths spiral flight as follows.
tion mechanism where a label indicates the cluster value of a
  vertice [15]. The label-based representation method is more
S Mi ; Fj ¼ Ri  ebt  cos ð2ptÞ þ Fj (3)
intuitive and easier to decode, which has lower computational
where S indicates the logarithmic spiral function. Ri is defined complexity than that of the locus-based representation mecha-
as the distance of the ith moth to the jth flame. b is a constant nism [43]. The moths and flames can be divided into two
for controlling the flight of moth. t denotes a random number groups through the MFO algorithm and their positions essen-
which ranges in ½1; 1. tially represent the network clustering. Thus, these two popu-
Ri is calculated by the following equation: lations adopt the same representation method. Specifically,
  assuming that a network boasts D vertices, the positions of the
Ri ¼ Fj  Mi  (4) ith moth and the ith flame are redefined as follows:
where Mi and Fj represents the ith moth and the jth flame, Mi ¼ fmi1 ; . . . ; miD g (6)
respectively.
In addition, to make the convergence of solutions more effi- where mij ; j 2 ½1; D represents the group value of the jth ver-
cient, MFO adopts an adaptive reduction strategy in the num- tice in the ith moth, which ranges between ½1; D as a random
ber of flames along with the iterative process. The adaptive integer. If xij ¼¼ xik , the jth vertice and the kth vertice of the
reduction process obeys the following formula: ith moth belong to the same cluster.
 
kðn  1Þ Fi ¼ ffi1 ; . . . ; fiD g (7)
nF ¼ round n  (5)
K
Similarly, fij ; j 2 ½1; D represents the group value of the
where the round operation means obtaining the nearest integer
jth vertice in the ith flame which ranges between ½1; D as a
of ðn  kðn1Þ
K Þ and the maximum number of flames is denoted random integer. If fij ¼¼ fik , it means that the jth vertice and
as n. k represents the current number of iteration and K indi-
the kth vertice of the ith flame belong to the same cluster.
cates the maximal iteration number. nF is denoted as the adap-
The discrete position of a moth or flame is illustrated in
tive number of flame.
Fig. 1, which gives an example of encoding and decoding
schemes. Fig. 1(a) shows a network topology with 8 nodes
III. THE PROPOSED DMFO METHOD
and can be divided into four clusters by encoding. Fig. 1(b)
In this section, the discrete processes of our DMFO method shows a decode representation, after clustering, the topology
for network clustering will be described in detail. First, network can be decoded into two cluster divisions. To help the

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
1780 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 9, NO. 3, MAY/JUNE 2022

Algorithm 1: The SFG process of DMFO method.

Input: Moths Mðk  1Þ, flames F ðk  1Þ, crossover probabil-
ity: pc, mutation probability: pm, max generation:
Max iter.
Output: Flames F ðkÞ.
1 while k 2 ½1; Max iter do
2 if k ¼ 1 then
3 foreach Mi of the Mð0Þ do
4 Calculate the modularity Q for each Mi ;
5 Sð0Þ Sort(Mð0Þ, descend) with the Q;
6 F ðkÞ ¼ Sð0Þ;
7 else
8 foreach Fi of the F ðk  1Þ do
9 if randð0; 1Þ < pc then
10 Randomly select one Fj from the neighboors;
11 Fi0 two-ways crossoverðFi ; Fj Þ;
12 Evaluate Fi0 ;
13 if Fi0 is better than Fi then Fig. 2. Illustration of two-way crossover. (a) The individual Pop1 and its net-
14 Fi ¼ Fi0 ; work structure. (b) The individual Pop2 and its network structure. (c) A ran-
dom two-way crossover of individuals Pop1 and Pop2 generates their children
15 if randð0; 1Þ < pm then New1 and New2. (d) The superior individual New1 and its network structure.
16 Fi00 NBM operator for Fi ;
17 Evaluate Fi0 ; are taken as the target individual and the source individual of
18 if Fi00 is better than Fi then
the crossover correspondingly. At last, two new descendants
19 Fi ¼ Fi00 ;
can be acquired by the two-way crossover at the same time,
20 Mðk  1Þ MergeðMðk  1Þ; F ðk  1ÞÞ;
21 foreach Mi of Mðk  1Þ do and the best individual Fi0 is selected to compare with Fi . If Fi0
22 Calculate modularity Q for each Mi ; is better than Fi , we substitute Fi with Fi0 , or vice versa.
23 Sðk  1Þ Sort(Mðk  1Þ, descend) with Q; The neighborhood-based two-way crossover is shown in
24 F ðkÞ Sðk  1Þf1 : ng; Fig. 2. Pop1 in Fig. 2(a) and Pop2 in Fig. 2(b) are taken as the
target individual and the source individual of the crossover
process, respectively. A node vi is first selected randomly in
order to find all the other nodes that have the same cluster
proposed DMFO converge more quickly, a label propagation identifier in Pop2. Then, we assign this cluster identifier to the
strategy [44] is leveraged, so that the random moth population same nodes in Pop1 and generate a new chromosome New1
initialization can be achieved. finally. The corresponding operations are performed in Pop1
to generate the new chromosome New2. In addition, when the
two-way crossover is performed, the neighborhood where
B. The SFG Process of DMFO
each moth pursues the crossover object is selected according to
In the original MFO, the SFG process is used to select the the distances among their aggregation multi-objective weight
best individuals of moths and flames so far as the next genera- vectors.
tion of flames, guiding the moths to find solutions in the search The NBM mutation operator is used in the SFG process
space. In our DMFO method, the SFG process is readjusted to inspired by [15]. The core idea of this process is as follows.
pursue the global optimum instead of the local optimum when First, a mutation probability pm is given. Then, a random
dividing networks. In detail, the best individuals are firstly number within the range [0, 1] for each vertice of each indi-
selected in the current two generations as flames to help moths vidual will be generated. Finally, if the number generated ran-
find the best solution. Considering the significance of the domly cannot surpass the mutation probability pm, its cluster
flames in the DMFO variant, the neighborhood-based two- value will be assigned to all its neighbor vertices.
way crossover and neighbor-based mutation (NBM) strate-
gies [44] are adopted in SFG process so that the high-perfor-
mance flames can be caught and prevent falling into the local C. The SFS Process of DMFO
optimum. Following that, the process combines the obtained In the original MFO, the SFS process is proposed to simu-
flames and the moths in this iteration, and the well-behaved late the behavior of the moth spiral flight to update its position.
individuals are chosen as the next-generation flames by the This updating strategy provides an excellent flight direction
modularity Q. The pseudo-code for the entire SFG process of for the moth to find a high-quality solution in the search space.
DMFO algorithm is shown in Algorithm 1. However, for the network clustering problem, in our proposed
The main idea of the neighborhood-based two-way cross- DMFO method, the SFS process is redesigned so that the effi-
over is following. As to the ith individual Fi ði ¼ 1; . . . ; nÞ in ciency and scope of the search can get optimized in a solution
the flames F , first, an individual Fj ðj ¼ 1; . . . ; HÞ will be space. In more detail, the distance U changes when updating a
selected randomly in its neighborhood. Afterward, Fi and Fj moth position based on (8).

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DISCRETE MOTH-FLAME OPTIMIZATION WITH AN L_2-NORM CONSTRAINT FOR NETWORK CLUSTERING 1781

Algorithm 2: The SFS process of DMFO method.

Input: Moths MðkÞ, flames F ðkÞ, max generation: Max iter,
moths or flames size: popsize.
Output: Moths Mðk þ 1Þ.
1 nF based on (5);
2 for i ¼ 1 : popsize do
3 if i < ¼ nF then
4 Ri Mi  Fi based on (9);
for j ¼ 1 : D do
6 b ¼ 1; t ¼ 1  2  randð0; 1Þ;
7 Uij Rij  Sigðjebt  cos 2ptjÞ based on (8);
8 if Uij ¼ 0 then
9 Mij ðk þ 1Þ Mij ðkÞ;
10 else
11 Mij ðk þ 1Þ Nbestij ;
12 if i > nF then
13 Ri Mi ðkÞ  FnF ðkÞ based on (9);
14 for j ¼ 1 : D do Fig. 3. Illustration of a moth using the redefined SFS process to update its
15 b ¼ 1; t ¼ 1  2  randð0; 1Þ; position. (a) The ith moth and its network structure. (b) The ith flame and its
16 Uij Rij  Sigðjebt  cos 2ptjÞ based on (8); network structure. (c) The regeneration process of the ith moth. (d) Newly
17 if Uij ¼ 0 then generated ith moth and its graph structure.
18 Mij ðk þ 1Þ Mij ðkÞ;
19 else follows:
20 Mij ðk þ 1Þ Nbestij ;

mkid if ud ¼ 0
id ¼
mkþ1
Nbestkd else
(13)
 
Ui ¼ Ri  Sig ebt  cos 2pt (8) where Nbestkd stands for the cluster value owned by the largest
part of neighbors of the dth vertice of a moth.
where b and t are the same as (3). The distance Ri between the Also, in the novel SFS process, the mechanism remains
moth Mi and the flame Fj is redefined as follows: unchanged that the number of flames automatically decreases,
Ri ¼ M i  F j (9) as shown in (5). In the newly designed DMFO algorithm,
Fig. 3 shows an example of a moth performing position update
where  is expressed as an XOR operator. If the correspond- based on a redefined SFS process, where ti denotes a set of
ing vertice cluster value between a moth Mi and a flame Fj is random numbers which ranges in [0,1]. The detailed algorithm
the same, it is 0, otherwise, 1. process of SFS process is given in Algorithm 2.
Suppose that Y ¼ ðy1 ; . . . ; yD Þ, X ¼ ðx1 ; . . . ; xD Þ. In (8),
the function Y ¼ SigðXÞ is defined as follows: D. The Tchebycheff Decomposition Method With a
 Constraint
1 if rand(0, 1) < sigmoidðxi Þ
yi ¼ i ¼ 1; . . . ; D (10) Our proposed DMFO method addresses the network cluster-
0 else
ing problem based on the multi-objective evolutionary algo-
where the signomid function [45] is represented as follows: rithm framework with decomposition (MOEA/D) [46]. In
MOEA/D, a Tchebycheff decomposition method decomposes
1 the multi-objective optimization problem into a set of scalar
sigmoidðxÞ ¼ (11)
1 þ ex optimization sub-problems, which can obtain excellent perfor-
mance and is widely used in MOP [9], [15], [20], [39]. How-
The redefined U function evaluates and determines whether ever, the solutions obtained by the original Tchebycheff
a moth requires update. Based on this, we redefine the updat- decomposition with the uniform weight vectors are usually
ing rule of moth position as follows: not uniformly distributed, and the geometric properties of the
sub-objective functions in these original Tchebycheff decom-
Mi ðk þ 1Þ ¼ Ui Mi ðkÞ (12) position methods have not been studied in detail. Ma et al.
[47] adopted a Tchebycheff decomposition method (p-Tch)
where k is the number of iterations. The operator is crucial with lp -norm constraints on the direction vector, in which the
in the moth updating rule. Choosing a good operator will sub-objective function has clear geometric properties, and the
improve the performance of the moth. Assuming that the kth construction of each sub-problem is based on an appropriate
generation moth Mi ðkÞ ¼ ðmki1 ; . . . ; mkiD Þ, the ðk þ 1Þth gen- direction vector instead of improving the weight vector w
eration moth is Mi ðk þ 1Þ ¼ ðmkþ1 i1 ; . . . ; miD Þ and Ui ¼
kþ1
used in Tchebycheff decomposition. Since different p values
ðu1 ; . . . ; uD Þ, the specific updating strategy of moth is as impose different competitive pressures on the sub-problems of

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
1782 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 9, NO. 3, MAY/JUNE 2022

Algorithm 3: The DMFO framework. First, the initialization phase is used to initialize moths and
Input: The adjacency matrix A of a network, max generation flames, which can be completed in the linear time OðjEjÞ.
Max iter, mutation probability: pm, neighborhood size: Then, the redesigned SFG process aims to select the flame,
niche, moths or flames size: popsize. which mainly uses the two-way crossover and NBM techni-
Output: A best cluster divisions C ¼ fCi gci¼1 of a given network. ques. Here, the time complexity of such two techniques are
1 Initialize Mð0Þ; both OðNÞ. In addition, the SFG takes OðN popsize
2 Generate a uniformly distributed weight vector w ¼ ðw1 ; . . . ; logðpopsizeÞÞ to select the optimal flames based on the modu-
wpopsize Þ; larity of descending order. Namely, the time complexity of the
3 foreach each moth and flame do SFG process is OðNþN popsize logðpopsizeÞÞ, where
4 Initialize the neighborhood based on the weight vectors; popsize is size of moths or flames. Moreover, the redesigned
5 Initialize reference point z .
SFS process updates moths by redesigning the moth flight
6 while k 2 ½1; Max iter do
strategy for specific problems. That is, the time complexity of
7 F ðkÞ execute Alg:1 : SFG process;
8 MðkÞ execute Alg:2 : SFS process; SFS process is OðD popsizeþNÞ. Finally, DMFO uses the
9 if randð0; 1Þ < pm then NBM operation again for the local search, which takes OðNÞ
10 Mi ðkÞ the NBM operator; time. Therefore, the overall time complexity of our proposed
11 Evaluate all moths MðkÞ based on (15); DMFO is OðMax iterðjEjþN popsize logðpopsizeÞÞ, whe-
12 if Mi ðkÞ is better than Mi ðk  1Þ then re Max iter is the max iteration number.
13 Mi ðkÞ Mi ðkÞ;
14 else
15 Mi ðkÞ Mi ðk  1Þ; IV. EXPERIMENTS
16 Update reference point z ; Aiming to test the behavior of our DMFO method, the exper-
17 return C ¼ fCi gci¼1 ; imental results and analyses are presented in this section.
Section IV-A introduces the experimental settings. Section IV-
B compares our proposed method with some state-of-the-art
MOEA/D, MOEA/D-2TCHMFI proves that choosing p = 2 approaches in both synthetic and real-world networks.
can improve the uniformity of the subproblem update regions, Section IV-C shows the contribution of an optimized Tch-
and the subproblem objective function in 2-Tch has a clearer ebycheff decomposition approach for network clustering.
geometric property in terms of Euclidean distance than that of Section IV-D focuses on the influence of the newly designed
using larger p values [47]. Therefore, the Tchebycheff decom- procedures, i.e., Algorithm 1 and Algorithm 2.
position using l2 -norm form constraint is an excellent decom-
position method. Based on the above decomposition method,
A. Experimental Design
the multi-objective optimization problem is transformed into a
single-objective sub-problem. This paper solves the uneven 1) Datasets and Parameter Setting: The extended Girvan-
distribution of the weight vector of the sub-problem in the pre- Newman (GN) benchmark networks [10] and real-world net-
vious decomposition algorithm that improves the overall works are selected to test the performance of our proposed
application performance of all sub-problem. Finally, we can DMFO method for network clustering.
achieve a better balance about the convergence and diversity As listed in Table I, there are seven real-world networks,
of the algorithm solutions. The definition of the 2-Tch method including the Zachary’s Karate Club network (G1 ) [48], the
is given in (14). Bottlenose Dolphin network (G2 ) [49], the American College
Football Network (G3 ) [10], the Books about US Politics net-
  work (G4 ) [50], the Santa Fe Institute (SFI) network (G5 ) [51],
fi ðxÞ  zi
min gptch ðFðxÞjw; z Þ ¼ max (14) and the Email network (G6 ) [52], and the Power grid network
x2V 1iq wi
(G7 ) [48]. In the proposed DMFO, all datasets cannot iterate
over 80 times, and the population scale is customized by the
where w ¼ ðw1 ; . . . ; wq Þ with kwk2 ¼ 1 and w1 ; . . . ; wq 0.
dataset size. In order to save computational cost in large data-
z ¼ ðz1 ; . . . ; zq Þ denotes the reference point (i.e., zi ¼ min
sets, the size of population for G1 and G2 is 150, and 100 for
fi ðxjx 2 V; i ¼ 1; . . . ; q) in which q represents the number of
other datasets. Besides that, DMFO updates moths mainly by
objective functions. flames, but it often fails to avoid the local optimization. In
view of this, the mutation rate and crossover rate are pro-
E. Computation Complexity Analysis
moted, i.e., pm ¼ 0:3 and pc ¼ 0:9, on the basis of their
The framework of our proposed DMFO method is presented corresponding references. Moreover, to reduce the statisti-
in Algorithm 3. Moreover, Fig. 4 illustrates the overall flow of cal error, all experiments are the averaged results over 10
DMFO method. As shown in Algorithm 3 and Fig. 4, the main running times.
loop of the overall DMFO is mainly composed of two rede- 2) Evaluation Metrics: In order to measure the perfor-
signed processes (i.e., the SFG and SFS processes). Given a mance of an algorithm, the modularity Q [10] and the normal-
network G with N vertices and jEj edges, the computational ized mutual information(NMI) [53] are usually used as
complexity of DMFO method are analyzed as follows. evaluation metrics. The modularity Q is defined in (15).

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DISCRETE MOTH-FLAME OPTIMIZATION WITH AN L_2-NORM CONSTRAINT FOR NETWORK CLUSTERING 1783

Fig. 4. The flow chart of the proposed DMFO method. It can be seen from the figure that the overall DMFO framework is mainly composed of two redesigned
processes (i.e. SFG and SFS processes).
Pc A Pc B  
2 Cij log Cij N=Ci: C:j
TABLE I
NMI ¼ PcA
i¼1 j¼1
PcB   (16)
i¼1 Ci: log ðCi: =N Þ þ
THE RELEVANT INFORMATION OF THE REAL-WORLD NETWORKS
j¼1 C:j log C:j =N

where cA ðcB Þ denotes the number of clusters in partition

AðBÞ, Ci: ðC:j Þ means the total elements of C in the ith row
(jth column), and N represents the total number of vertices of
a network. The value of NMIðA; BÞ is between 0 and 1. If
NMIðA; BÞ ¼ 1, it means that A and B are exactly the same;
if NMIðA; BÞ ¼ 0, it means that A and B are completely
different.
The comparison is made between the performance of the
proposed DMFO and eleven new or popular network cluster-
ing algorithms, namely, GA-Net [13], MOGA-Net [20],
MODPSO [15], MODBSA/D [39], QDM-PSO [16], MOPSO-

X   ! Net [17], DCRO [14], DMFWA [12], RMOEA [9], MOCD-

c
ls ds 2 ACO [18], and DIM-MOEA/D [19].
Q¼  (15)
s¼1
2jEj 2jEj
B. Experimental Results
where c denotes how many clusters a network owns; jEj repre- 1) Experiments on Synthetic Networks: In this subsection,
sents the total number of edges in the network; ls denotes the the well-known synthetic test networks, GN-benchmark net-
total number of edges among vertices in a cluster s, and ds works [10], are generated to verify the ability of our proposed
means the total degree of vertices in s. When estimating the per- algorithm. The networks contain 128 vertices, which are
formance of network clustering algorithms, a larger Q tends equally divided into four clusters with 32 vertices in each clus-
to indicate a higher clustering quality of an algorithm for a ter, and the average degree of each vertice is 16. m denotes a
network. More specifically, a smaller difference between mixing parameter. Each vertice shares 1  m percentage edges
Q and 1 tends to indicate a clearer community structure of with other vertices in the same cluster and the remaining m
a network. edges is shared with the other vertices of a network. In gen-
The normalized mutual information NMIðA; BÞ can be for- eral, a higher m means the community structure blurs more
mulized as follows: seriously. When the mixing parameter m < 0:5, the neighbors

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
1784 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 9, NO. 3, MAY/JUNE 2022

Fig. 5. The NMI values obtained by running 10 times on GN extended Fig. 6. The boxplot of the statistic values with modularity (Q) obtained by
benchmark datasets for different algorithms. The figure indicates the superior DMFO on the GN extended benchmark networks. It can be seen from the
performance of our DMFO than other methods on synthetic datasets. figure that our DMFO method has better robustness in network clustering.

TABLE II
of each vertice in the same cluster surpasses that of other RESULTS OBTAINED BY THE COMPARED ALGORITHMS ON G1 . THE RESULTS
neighbors in other clusters, and the networks have a clear SHOW THAT DMFO CAN FIND THE REAL COMMUNITY STRUCTURE AND ALSO
GET THE BEST MODULARITY Q VALUE
community structure. When m > 0:5, the community struc-
ture will become exceedingly blurred. Therefore, the mixing
parameters m ranges from 0.0 to 0.5, with a spacing of 0.05,
giving rise to a total of 11 synthetic networks in our experi-
ments. We test the proposed algorithm on this set of networks
and utilize NMI and Q value as criteria to estimate the parti-
tioning performance.
Fig. 5 provides the statistical results of the NMI over 10
runs for each algorithm with varying mixing parameters from
0 to 0.5. As can be seen, the proposed DMFO algorithms can
detect the network with the mixing parameter m being 0.45,
and the rest of the algorithms cannot detect it. Among them,
GA and MOGA-Net algorithms have the worst performance,
whose detection scope is restricted to when m is below 0.15.
When m > 0:35, GA-Net completely fails. For the MODPSO,
QDMPSO, MODBSA/D, and MOPSO-Net algorithms, the
performance is similar, and the real division of the network
community below 0.45 can be detected. To summarize, our
proposed DMFO algorithm surpasses the extended GN bench-
mark networks in performance. the administrator and the instructor in the period. Table II
Fig. 6 plots statistical results of Q obtained by DMFO algo- shows the experimental results for Q and NMI of different
rithm on the extended GN-benchmark datasets. As can be algorithms. Results show that the average NMI of MODPSO,
seen from Fig. 6, as the mixing parameter m gradually MODBSA/D, MOPSO-Net, DMFWA, MOCD-ACO, DIM-
increases from 0 to 0.5, the Q value obtained by the algorithm MOEA/D, and DMFO algorithm is 1, which means that real
decreases linearly. Overall, the robustness of the DMFO algo- clusters can be mined by them. The GA-Net, MOGA-Net, and
rithm is very strong, especially when m < 0:4, the standard DCRO methods cannot achieve the accurate division. It can
deviation of the Q value is 0. be seen from this network that our DMFO is effective and
2) Experiments on Real-World Networks: In this subsec- robust.
tion, we estimate the network clustering performance of G2 is an animal social network consisting of 62 bottlenose
DMFO on seven popular real-world networks. The results dolphins living off Doubtful Sound, New Zealand over 7
obtained by our proposed DMFO are represented in boldface. years, based on the social behavior of dolphins [49]. G2 is nat-
G1 is a social network of friendships among 34 members of urally divided into two large clusters with gender(i.e., the
a karate club spanning 2 years [48]. The club is divided into female and the male). In Table III, the results of different
two groups of the same size due to the inconsistence between methods on G2 are reported. As it can be seen from the results,

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DISCRETE MOTH-FLAME OPTIMIZATION WITH AN L_2-NORM CONSTRAINT FOR NETWORK CLUSTERING 1785

TABLE III TABLE V

RESULTS OBTAINED BY THE COMPARED ALGORITHMS ON G2 . WE CAN CON- RESULTS OBTAINED BY THE COMPARED ALGORITHMS ON G4 . THE RESULTS
CLUDE THAT THE REAL COMMUNITY STRUCTURE AND THE BEST MODULAR- VERIFY THAT THE BEST MODULARITY Q AND NMI VALUE CAN BE
ITY Q VALUE CAN BE OBTAINED BY DMFO OBTAINED BY DMFO, WHICH CAN OBTAIN SOLUTIONS THAT ARE COMPETI-
TIVE WITH OTHER ALGORITHMS IN TERMS OF BOTH Q AND NMI

TABLE IV G3 . It can be observed from Table IV that none of algorithms

RESULTS OBTAINED BY THE COMPARED ALGORITHMS ON G3 . THE DMFO
CAN ACQUIRE THE OPTIMAL MODULARITY Q VALUE AND THE BEST NMI can mine the true clustering structure of such network. How-
VALUE, WHICH PROVES THAT THE DMFO IS BETTER THAN SEVERAL STATE- ever, DMFO can obtain the best NMI value and Q value. The
OF-THE-ART METHODS stability can also be detected from the standard deviation
equalling 0 for DMFO. It can be concluded from this example
that the proposed DMFO algorithm is very promising and can
obtain outstanding solutions.
G4 consists of 105 vertices and 441 edges indicating the
books about American politics and the relations of two books
about American politics that are frequently co-purchased by
the same buyers [51]. Table V reports the results of different
algorithms with respect to G4 . As can be seen, none of algo-
rithms can detect the true clustering structure of G4 . However,
DMFO can obtain the optimal maximal Q and NMI value,
which are 0.5272 and 0.6501, respectively. At the same time,
DMFO can also obtain the optimal average value of Q and
NMI, which are 0.5268 and 0.6275 respectively. In general,
the proposed DMFO algorithm has better performance than
other methods for network clustering in G4 .
For three unknown divisions networks, we estimate the per-
formance of different algorithms by comparing the modularity
Q. First, G5 consists of 271 vertices which represent scientists
MODPSO, MODBSA/D, MOPSO-Net, RMOEA, MOCD- in the residence and edges indicates that a pair of scientists
ACO and DMFO can find the true cluster successfully (their have collaborated on one or more articles in the same
average NMI is 1). Furthermore, in addition to the GA-Net, period [50]. We only do experiments with the largest compo-
DMFWA, and DCRO, the remaining algorithms can find the nent of G5 with 118 vertices in this part. Then, G6 is an email
real network structure in the best state (i.e., the maximum network consisting of 1133 vertices representing users and
NMI is 1). The proposed DMFO method achieves the highest 5451 edges representing communications between different
value in terms of the modularity Q, which proves our proposed users [52]. Finally, the large-scale G7 simulates the topology
method is outstanding. From the perspective of both Q and of the high-voltage power grid in the Western States of Amer-
NMI, our DMFO is still quite promising and effective. ica [48]. The vertices represent the power base stations and
G3 is divided into 12 clusters consisting of 115 vertices and the edges denote the high-voltage transforming line between
616 edges which stand for the football groups and the regular- two stations. Table VI shows the experimental results of dif-
season matches between different groups, correspond- ferent algorithms for G5 , G6 and G7 network. As can be seen
ingly [50]. Table IV shows the results for all algorithms on from the Table VI, DMFO algorithm can obtain the best

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
1786 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 9, NO. 3, MAY/JUNE 2022

TABLE VI
RESULTS OBTAINED BY THE COMPARED ALGORITHMS ON G5 , G6 AND G7 IN TERMS OF THE MODULARITY Q. IT CAN BE SEEN FROM THE TABLE THAT THE PRO-
POSED DMFO METHOD CAN OBTAIN THE OPTIMAL MODULARITY AMONG ALL COMPARISON ALGORITHMS

TABLE VII can calculate the optimal average value of Q successfully on four
COMPARISON RESULTS IN Q AND NMI ON THE FOUR REAL-WORD DATA- real-world networks. Moreover, DMFO and DMFO-origin can
SETS. THE TABLE INDICATES THAT OUR DMFO METHOD USING A TCHEBY- obtain NMI=1 on both G1 and G2 networks, which verifies the
CHEFF DECOMPOSITION METHOD WITH AN l2 -NORM CONSTRAINTS CAN
OBTAIN FURTHER IMPROVED NETWORK CLUSTERING RESULTS THAN THE high clustering accuracy of the proposed method. Additionally,
DMFO-ORIGIN USING AN ORIGINAL DECOMPOSITION METHOD our DMFO using an improved Tchebycheff decomposition
method with an l2 -norm constraint on direction vectors can effec-
tively find the optimal cluster partition results compared to an
original Tchebycheff method.

D. Influence of the Newly Designed Procedures

In this section, we implement two variants of MFO (i.e.,
MFO-SFG and MFO-SFS) to analyze the performance of the
newly designed procedures. More specifically, the main differ-
ence between two MFO variants and DMFO is the implemen-
tation of SFS and SFG processes. MFO-SFG uses the process
proposed in Algorithm 1, whereas MFO-SFS uses the process
modularity Q value of 0.7506, 0.5308 and 0.8423 in G5 , G6 described in Algorithm 2.
and G7 , respectively. In general, DMFO method exhibits The networks used in this section are extended GN benchmark
strong competitiveness over all other algorithms in the case of networks, which are based on the original GN network [10]. An
real network with unknown partitioning. extended GN benchmark consists of 128 vertices with an average
degree of 16. The vertices are divided into four communities,
each of which has 32 vertices. Each vertex shares a fraction m of
C. Contribution of an l2 -Norm Constraint on Direction Vectors
its edges with vertices from the other three communities, and a
In this section, we select G1 , G2 , G3 , and G4 networks to ver- fraction 1m with the vertices from its own community. We call
ify the performance of an improved Tchebycheff decomposition m2[0,1] the mixing parameter. When the mixing parameter
method to decompose multi-objective optimization problems. m < 0.5, the neighbors of each vertice in the same cluster surpass
Specifically, the original Tchebycheff method is used in DMFO that of other neighbors in other clusters, and a network has a clear
to replace the improved method to decompose MOP as a com- community structure. When m > 0.5, the community structure
parison algorithm. Here, this comparison algorithm is denoted will become exceedingly blurred. Therefore, the mixing parame-
as DMFO-origin. In addition, except for using different decom- ters m ranges from 0 to 0.5, with a spacing of 0.05, giving rise to
position methods in two algorithms, the rest of parameters are a total of 11 synthetic networks in our experiments. For each net-
set the same. All experiments are repeated 10 times indepen- work, we ran all three algorithms once and recorded the NMI and
dently to acquire the average value. modularity. Fig. 7 shows the results, and each point corresponds
Table VII shows the experimental results of DMFO and to an average of 10 runs. All algorithms use the same parameter
DMFO-origin for four networks in terms of Q and NMI. The configuration.
proposed DMFO method and the best results are marked using It can be seen from Fig. 7 that when the mixing parameter m
bold font. Judging from Table VII, the proposed DMFO method is small (m  0.15), all three algorithms can find the true

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DISCRETE MOTH-FLAME OPTIMIZATION WITH AN L_2-NORM CONSTRAINT FOR NETWORK CLUSTERING 1787

to some more promising solution areas and ensure the superi-

ority of the algorithm performance.
After comparing the performance of DMFO with that of
MFO-SFS, we can conclude that DMFO surpasses MFO-SFS
in both the average NMI and the average Q. The only differ-
ence is that DMFO employs both Algorithm 1 and Algorithm
2, whereas MFO-SFS only uses Algorithm 2. The difference
makes DMFO achieve better performance than MFO-SFS.
Fig. 7. The calculation results of (a) Modularity and (b) NMI about the three
The improved SFS strategy in Algorithm 2 provides the ability
MFO variants on the expanded GN network. DMFO algorithm obtains the bet- of both DMFO and MFO-SFS to jump from one area in the
ter results. search space to another area, and gradually move to the prom-
ising areas. But it is impossible for the program in Algorithm
TABLE VIII 2 to thoroughly scan the promising area, this defect hinders
WILCOXON SIGNED RANKS TEST RESULTS. DMFO SHOWS AN IMPROVEMENT
OVER MFO-SFS AND MFO-SFG, WITH A LEVEL OF SIGNIFICANCE a ¼ 0:05. the further improvement of MFO-SFS. However, based on
IN ADDITION, MFO-SFS ALSO SHOWS A SIGNIFICANT IMPROVEMENT OVER Algorithm 1, DMFO can overcome the defect of Algorithm 2,
MFO-SFG, WITH A LEVEL OF SIGNIFICANCE a ¼ 0:05 which further explains the better performance of DMFO.

V. CONCLUSION
We propose a decomposition-based multi-objective discrete
moth-flame optimization method (DMFO) to optimize net-
work clustering. In detail, what is novel in the proposed
method is that a good tradeoff has been achieved between
division, that is, NMI is equal to 1. When m is greater than exploration and exploitation when clustering networks by
0.15, it is difficult for MFO-SFG to find the real partition. redefining the two processes of MFO algorithm, which ensures
Both Q and NMI values of MFO-SFG dropped rapidly. How- the better clustering results. Additionally, the improved Tche-
ever, the performance of MFO-SFS and DMFO has not bycheff decomposition method (2-Tch) divides and conquers
changed much. When m is equal to 0.3, two methods can still the network clustering problem in our algorithm. Experimen-
find close to the real community division. When m¼0.35, the tal results based on synthetic and real-world datasets can prove
performance of MFO-SFS begins to degrade. The average Q the superiority of DMFO method than other comparison algo-
value of MFO-SFS falls below 0.35, and the average NMI rithms. Moreover, the Tchebycheff method with an l2 -norm
falls below 0.53. Conversely, DMFO performs well even constraint on direction vectors can obtain better clustering
when m is 0.45, specifically, its average NMI is equal to 1 results than the original method.
and its average Q is greater than 0.51. There still remains some work related with DMFO that
In order to further study the performance differences of deserves to be further investigated. The proposed DMFO has
three algorithms, we perform the Wilcoxon signed rank shown the effectiveness of detecting single-layer network
test [54] on the results obtained by all three algorithms. The clustering. It is interesting to extend it to community detection
results are listed in Table VIII. According to the p value, it in multi-layer complex networks, because most real-world
can be seen that DMFO shows a significant improvement over communities are multi-layered. Moreover, it is worth noting
the MFO-SFS and MFO-SFG algorithms, with a level of sig- that the computational efficiency of ultra-large-scale networks
nificance a¼0.05. While the MFO-SFS algorithm shows a sig- algorithms used for community detection is usually low.
nificant improvement over the MFO-SFG algorithm, with Therefore, in the future work, how to design an optimized var-
a¼0:05. In other words, DMFO performs the best among the iant of moth-flame to effectively reduce the computational
three algorithms. Compared with MFO-SFG, DMFO can efficiency of the algorithm without affecting the accuracy of
obtain better NMI and better modularity. This result shows clustering will be the focus of research.
that although the chromosomes generated by the crossover
and mutation strategies in Algorithm 1 can expand the new REFERENCES
search space, the population begins to converge and the struc- [1] F. Chiti and C. M. Dobson, “Protein misfolding, amyloid formation, and
ture of chromosomes begins to become similar after constant human disease: A summary of progress over the last decade,” Annu.
Rev. Biochem., vol. 86, no. 1, pp. 27–68, Jun. 2017.
calculation, which cause the chromosomes generated by cross- [2] C. Gao and J. Liu, “Network-based modeling for characterizing human
over operation are still similar. Whats more, in the late stage collective behaviors during extreme events,” IEEE Trans. Syst., Man,
of the calculation, due to the high fitness of the entire popula- Cybern., Syst., vol. 47, no. 1, pp. 171–183, Jan. 2017.
[3] E. Strano, M. P. Viana, A. Sorichetta, and A. J. Tatem, “Mapping road
tion, the chromosomes generated by the mutation are less network communities for guiding disease surveillance and control strat-
likely to survive the selection. Therefore, in general, although egies,” Sci. Rep., vol. 8, no. 1, Mar. 2018, Art. no. 4744.
the strategy of Algorithm 1 can prevent the algorithm from [4] E. Stalidzans, I. Mozga, J. Sulins, and P. Zikmanis, “Search for a mini-
mal set of parameters by assessing the total optimization potential for a
falling into the local optimum quickly, we still need to dynamic model of a biochemical network,” IEEE-ACM Trans. Comput.
improve the diversity of algorithm, so as to guide Algorithm 1 Biol. Bioinf., vol. 14, no. 4, pp. 978–985, Jul./Aug. 2017.

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.
1788 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 9, NO. 3, MAY/JUNE 2022

[5] K. He, Y. Li, S. Soundarajan, and J. E. Hopcroft, “Hidden community [31] Y. Xu and H. Chen, “An efficient chaotic mutative moth-flame-inspired
detection in social networks,” Inf. Sci., vol. 425, pp. 92–106, Jan. 2018. optimizer for global optimization tasks,” Expert Syst. Appl., vol. 192,
[6] L. Sun, J. G. Jin, K. W. Axhausen, D.-H. Lee, and M. Cebrian, pp. 135–155, Sep. 2019.
“Quantifying long-term evolution of intra-urban spatial interactions,” J. [32] Z. T. Li, J. Liu, and K. Wu, “A multiobjective evolutionary algorithm
Roy. Soc. Interface, vol. 12, no. 102, Jan. 2015, Art. no. 20141089. based on structural and attribute similarities for community detection in
[7] F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and D. Parisi, attributed networks,” IEEE Trans. Cybern., vol. 48, no. 7, pp. 1963–1976,
“Defining and identifying communities in networks,” Proc. Nat. Acad. Jul. 2018.
Sci. USA, vol. 101, no. 9, pp. 2658–2663, Feb. 2004. [33] C. Gao, Z. Chen, X. Li, and Z. Wang, “Multiobjective discrete particle
[8] M. E. J. Newman, “Fast algorithm for detecting community structure in swarm optimization for community detection in dynamic networks,”
networks,” Phys. Rev. E, vol. 69, no. 6, Jun. 2004, Art. no. 066133. Europhysics Lett., vol. 122, no. 2, Apr. 2018, Art. no. 28001.
[9] X. Zhang, K. Zhou, H. Pan, L. Zhang, X. Zeng, and Y. Jin, “A network [34] C. Gao, M. X. Liang, and X. H. Li, “Network community detection
reduction-based multiobjective evolutionary algorithm for community based on the Physarum-inspired computational framework,” IEEE-ACM
detection in large-scale complex networks,” IEEE Trans. Cybern., Trans. Comput. Biol. Bioinform., vol. 15, no. 6, pp. 1916–1928, Nov.
vol. 50, no. 2, pp. 703–716, Feb. 2020. 2018.
[10] M. Girvan and M. E. J. Newman, “Community structure in social and [35] C. Pizzuti and A. Socievole, “Multiobjective optimization and local
biological networks,” Proc. Nat. Acad. Sci. USA, vol. 99, no. 12, merge for clustering attributed graphs,” IEEE Trans. Cybern., vol. 50,
pp. 7821–7826, Jun. 2002. no. 12, pp. 4997–5009, Dec. 2020.
[11] B. K. Patra, S. Nandi, and P. Viswanath, “A distance based clustering [36] Z. Li and J. Liu, “A multi-agent genetic algorithm for community detec-
method for arbitrary shaped clusters in large datasets,” Pattern Recog- tion in complex networks,” Phys A. Statist. Mech. Appl., vol. 499, no. 2,
nit., vol. 44, no. 12, pp. 2862–2870, Dec. 2011. pp. 336–347, 2016.
[12] M. Guendouz, A. Amine, and R. M. Hamou, “A discrete modified fire- [37] M. Gong, L. Ma, Q. Zhang, and L. Jiao, “Community detection in net-
works algorithm for community detection in complex networks,” Appl. works by using multiobjective evolutionary algorithm with decom-
Intell., vol. 46, no. 2, pp. 373–385, Mar. 2017. position,” Phys. A, vol. 391, no. 15, pp. 4050–4060, Aug. 2012.
[13] C. Pizzuti, “GA-Net: A genetic algorithm for community detection in [38] R. Shang, S. Luo, W. Zhang, R. Stolkin, and L. Jiao, “A multiobjective
social networks,” in Proc. 10th Int. Conf. Parallel Problem Solving evolutionary algorithm to find community structures based on affinity
Nature, 2008, pp. 1081–1090. propagation?,” Phys. A, vol. 453, pp. 203–227, Jul. 2016.
[14] H. H. Chang, Z. R. Feng, and Z. G. Ren, “Community detection using [39] F. Zou, D. Chen, S. Li, R. Lu, and M. Lin, “Community detection in
dual-representation chemical reaction optimization,” IEEE Trans. complex networks: Multi-objective discrete backtracking search optimi-
Cybern., vol. 47, no. 12, pp. 4328–4341, Dec. 2017. zation algorithm with decomposition,” Appl. Soft. Comput., vol. 53,
[15] M. Gong, Q. Cai, X. Chen, and L. Ma, “Complex network clustering by pp. 285–295, Apr. 2017.
multiobjective discrete particle swarm optimization based on decom- [40] M. A. E. Aziz, A. A. Ewees, and A. E. Hassanien, “Whale optimization
position,” IEEE Trans. Evol. Comput., vol. 18, no. 1, pp. 82–97, algorithm and moth-flame optimization for multilevel thresholding
Feb. 2014. image segmentation,” Expert Syst. Appl., vol. 83, pp. 242–256, 2017.
[16] L. Li, L. Jiao, J. Zhao, R. Shang, and M. Gong, “Quantum-behaved dis- [41] C. Li, S. Li, and Y. Liu, “A least squares support vector machine model
crete multi-objective particle swarm optimization for complex network optimized by moth-flame optimization algorithm for annual power load
clustering,” Pattern Recognit., vol. 63, pp. 1–14, Mar. 2017. forecasting,” Appl. Intell., vol. 45, no. 4, pp. 1166–1178, 2016.
[17] S. Rahimi, A. Abdollahpouri, and P. Moradi, “A multi-objective particle [42] M. Wang et al., “Toward an optimal kernel extreme learning machine
swarm optimization algorithm for community detection in complex using a chaotic moth-flame optimization strategy with applications in
networks,” Swarm Evol. Comput., vol. 39, pp. 297–309, Apr. 2018. medical diagnoses,” Neurocomputing, vol. 267, no. 6, pp. 69–84, 2017.
[18] P. Ji, S. Zhang, and Z. Zhou, “A decomposition-based ant colony opti- [43] J. Handl and J. Knowles, “An evolutionary approach to multi-objective
mization algorithm for the multi-objective community detection,” J. clustering,” IEEE Trans. Evol. Comput., vol. 11, no. 1, pp. 56–76,
Ambient Intell. Humanized Comput., vol. 11, no. 1, pp. 173–188, Feb. 2007.
Jan. 2020. [44] M. Gong, Q. Cai, Y. Li, and J. Ma, “An improved memetic algorithm for
[19] F. Zou, D. Chen, D. Huang, R. Lu, and X. Wang, “Inverse modelling- community detection in complex networks,” in Proc. IEEE Congr. Evol.
based multi-objective evolutionary algorithm with decomposition for Comput., 2012, pp. 1–8.
community detection in complex networks,” Phys. A, vol. 513, pp. 662– [45] J. Kennedy and R. C. Eberhart, “A discrete binary version of the particle
674, Jan. 2019. swarm algorithm,” in Proc. IEEE Int. Conf. Syst., Man, Cybern.1997,
[20] C. Pizzuti, “A multiobjective genetic algorithm to find communities in pp. 4104–4108.
complex networks,” IEEE Trans. Evol. Comput., vol. 16, no. 3, pp. 418– [46] Q. Zhang and H. Li, “MOEA/D: A multi-objective evolutionary algo-
430, Jun. 2012. rithm based on decomposition,” IEEE Trans. Evol. Comput., vol. 11,
[21] X. Liu, F. Zhang, X. Li, C. Gao, and J. Liu, “Multi-objective discrete no. 6, pp. 712–731, Dec. 2007.
moth-flame optimization for complex network clustering,” in Proc. 25th [47] X. Ma, Q. Zhang, G. Tian, J. Yang, and Z. Zhu, “On Tchebycheff
Int. Symp. Methodologies Intell. Syst., 2020, pp. 372–382. decomposition approaches for multi-objective evolutionary opti-
[22] G. Wang and S. Deb, “Monarch butterfly optimization,” Neural Comput. mization,” IEEE Trans. Evol. Comput., vol. 22, no. 2, pp. 226–244,
Appl., vol. 31, pp. 1995–2014, 2019. Apr. 2018.
[23] A. Heidari and S. Mirjalili, “Harris hawks optimization: Algorithm and [48] J. Duch and A. Arenas, “Community detection in complex networks
applications,” Future Gener. Comput. Syst., vol. 97, pp. 849–872, 2019. using extremal optimization,” Phys. Rev. E, vol. 72, no. 2, Aug. 2005,
[24] S. Li and H. Chen, “Slime mould algorithm: A new method for stochastic Art. no. 027104.
optimization,” Future Gener. Comput. Syst., vol. 111, pp. 300–323, 2020. [49] D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase, E. Slooten, and
[25] Y. Yang and H. Chen, “Hunger games search: Visions, conception, S. M. Dawson, “The bottlenose dolphin community of doubtful sound
implementation, deep analysis, perspectives, and towards performance features a large proportion of long-lasting associations,” Behav. Ecol.
shifts,” Expert Syst. with Appl., vol. 177, 2021, Art. no. 114864. Sociobiol., vol. 54, no. 4, pp. 396–405, Sep. 2003.
[26] G. Wang, “Moth search algorithm: A bio-inspired metaheuristic algo- [50] M. Girvan and M. E. J. Newman, “Community structure in social and bio-
rithm for global optimization problems,” Memetic Comput., vol. 10, logical networks,” in Proc. Nat. Acad. Sci. USA, 2002, pp. 7821–7826.
no. 2, pp. 151–164, Jun. 2018. [51] L. N. F. Ana and A. K. Jain, “Robust data clustering,” in Proc. IEEE
[27] S. Mirjalili, “Moth-flame optimization algorithm: A novel nature- Comput. Soc. Conf. Comput. Vis. Pattern Recognit., 2003, pp. 128–133.
inspired heuristic paradigm,” Knowl.-Based Syst., vol. 89, pp. 228–249, [52] R. Guimer, L. Danon, A. Daz-Guilera, F. Giralt, and A. Arenas, “Self-
Nov. 2015. similar community structure in a network of human interactions,” Phys.
[28] C. Li, Z. Niu, Z. Song, B. Li, J. Fan, and P. X. Liu, “A double evolution- Rev. E, vol. 68, no. 6, Dec. 2003.
ary learning moth-flame optimization for real-parameter global optimi- [53] L. Danon, A. Dıaz-Guilera, J. Duch, and A. Arenas, “Comparing com-
zation problems,” IEEE Access, vol. 6, pp. 76700–76727, 2018. munity structure identification,” J. Statist. Mech., vol. 2005, no. 09,
[29] Y. Xu and H. Chen, “Enhanced moth-flame optimizer with mutation strat- pp. P09008–P09008, Sep. 2005.
egy for global optimization,” Inf. Sci., vol. 492, pp. 181–203, Aug. 2019. [54] J. Derrac, S. Garca, D. Molina, and F. Herrera, “A practical tutorial on
[30] W. Shan and Z. Qian, “Double adaptive weights for stabilization of the use of nonparametric statistical tests as a methodology for comparing
moth flame optimizer: Balance analysis, engineering cases, and medical evolutionary and swarm intelligence algorithms,” Swarm Evol. Comput.,
diagnosis,” Knowl.-Based Syst., vol. 214, Feb. 2021, Art. no. 106728. vol. 1, no. 1, pp. 3–18, 2011.

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on November 08,2023 at 08:42:16 UTC from IEEE Xplore. Restrictions apply.