See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/380541748

Community Detection for Information Propagation Relying on Particle

Competition

Conference Paper · August 2020

DOI: 10.1109/ICCC49849.2020.9238981

CITATION READS
1 8

5 authors, including:

Wenzheng Li Jingjing Wang

China People's Public Security University Beihang University
6 PUBLICATIONS 42 CITATIONS 270 PUBLICATIONS 8,855 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Wenzheng Li on 14 May 2024.

The user has requested enhancement of the downloaded file.

 Community Detection for Information Propagation
Relying on Particle Competition
Wenzheng Li∗ , Jingjing Wang† , Yong Ren†‡ , Dechun Yin∗ , Yijun Gu∗
∗
School of Information Network Security, People’s Public Security University of China, Beijing, 100038, China
†
Department of Electronic Engineering, Tsinghua University, Beijing, 100084, China
‡
Peng Cheng Laboratory, Shenzhen, 518055, China
Email: [email protected]

Abstract—In the process of information propagation, different propagation process of network changes, which only affects a
communities may be formed due to different opinions, interests small part of the network every time. In 2018, Agarwal et al.
or hobbies. However, for the application for information propa- proposed Dyprem algorithm based on the permanence index
gation, targeted dynamic community detection methods have not
been proposed previously. In this paper, we propose a particle to adjust the community structure of incremental nodes and
competition aided community detection scheme for the sake of maximize the value of the permanence index at each time
solving the dynamic community detection for information propa- point to ensure the quality of dynamic community detection
gation. In comparison to traditional particle competition models, [6]. In 2019, DynaMo algorithm [7] was proposed to complete
the particles in our proposed model are capable of performing the dynamic community detection by maximizing the modularity.
operations of walking, splitting and jumping and the domination
matrix of the network changes continuously. Moreover, with the In conclusion, extensive global search is often required in
aid of combining the previous particle competition experiences the current dynamic community detection. However, in re-
as well as the defined particle’s walking rules, our proposed cent years, the individual-based community detection method
community detection scheme can automatically select and update represented by particle competition mechanism [8] has been
the core nodes based on the results of previous evolution. Finally, gradually applied to complex networks.
simulation results show both the effectiveness and superiority of
our proposed particle competition aided community detection Competitive learning is an important machine learning
model for information propagation, which may have compelling method which has been widely used in semi-supervised learn-
applications in the context of the spread of opinions and computer ing [9] of complex network [10]. In the beginning, M. G.
viruses, etc. Quiles, Zhao Liang [11] proposed a community detection
Index Terms—information propagation, particle competition, method based on particle competition mechanism. They stud-
dynamic community detection, complex network.
ied the randomness and certainty of particles in complex net-
works. In their research, particles randomly walk and compete
I. I NTRODUCTION
with each other with established rules. Then, the function of
Information propagation in social networks has the char- community detection can be achieved by occupying nodes.
acteristics of diversified communication subjects, diversified In 2012, the model named Stochastic Competitive Learning
information views and prominent opinion leaders. Due to the in Complex Networks [12] was proposed. The function of
uniqueness of information propagation in social networks, it is dynamic community detection can be realized by that model.
necessary to establish a targeted dynamic community detection In addition, that model is a kind of non-linear stochastic
scheme for information propagation. dynamic mechanism with unsupervised competition, which
In the field of dynamic community detection, two main plays a driving role in the field of complex network systems.
methods for dynamic community detection have been pro- In 2018, LCU, a particle competition mechanism to solve
posed. The first kind of dynamic community detection methods the problem of semi-supervised learning in complex networks
was based on slice processing [1], while the other relies was proposed [13]. Particles randomly walk in the network
on dynamic increments [2]. To elaborate a little further, the according to the established rules until they are absorbed to
community detection method based on slice processing uses finish community detection. The LCU model of absorption and
time slices as a unit for classification, similarity comparison recovery behavior was introduced by semi-supervised learning
or gap comparison. The method of dynamic increment can to overcome the problem of slow convergence rate in the
determine the change of the previous community by the previous model [12] and the model is also well-performed
change of nodes or edges and can refine the evolution process in terms of accuracy at that time.
of each community at the same time. In 2016, Meng et al. Based on the existing research on dynamic community
proposed IDBL-INK algorithm [3] based on DBLINK [4] detection and competition mechanism, we believe that it is
to conduct incremental community detection. In 2017, Han feasible to propose a particle competition aided community
et al. proposed the ALPA algorithm [5]. This method takes detection model for the sake of solving the dynamic commu-
the historical information of the community into account and nity detection for information propagation. Specifically, the
updates community detection result based on the local label main contributions of this paper are summarized as follows.
 • Inspired by the information propagation in the real world number of particles in community c is denoted as k and the set
[14], we propose a particle competition aided community of particles in community c is recorded as Pc = {p1c , ..., pkc }.
detection model for information propagation by imitating
the behavior of information propagation in the network B. Particle Walking
[15] [16]. In the particle walking phase, each particle can perform a
• In order to adapt to the dynamic process of informa- walk, split or jump and the influence of particle walking on
tion propagation, particles perform operations such as nodes can be recorded by the network domination matrix. The
walking, splitting and jumping in the network of the domination matrix of the network can be given by:
proposed community detection model and change the 4
domination matrix of the network continuously. More- N (t) = [N1 (t), N2 (t), ..., Ni (t)]T , (2)
over, the proposed model can be applied to dynamic where Ni (t) is the domination vector of the node vi . In the
community detection by automatically changing the core particle walking phase, particles can select adjacent nodes
of the community each time according to the domination randomly as targets to walk. The node domination vector of
matrix. node vi at time t can be formulated as:
• We have collected three groups of tweets on Twitter,
∆
including the Manchester terrorist attack, Parson Green Ni (t) = [Ni1 (t), Ni2 (t), ..., Nic (t)]T . (3)
subway station explosion and COVID-19 to build three
The edge domination vector between nodes vi and vj at
information propagation networks. Based on those net-
time t can be formulated as:
works, both the effectiveness and superiority of our
∆
1 2 c
proposed particle competition aided community detection Eij (t) = [Eij (t), Eij (t), ..., Eij (t)]T . (4)
model for information propagation can be found with
the comparison of other dynamic community detection Suppose Ñic (t) denotes the number of particles in commu-
methods proposed in recent years. nity c on node vi at time t. Besides, domination vectors can
The rest of the paper is organized as follows. Section be get by equations:
II introduces the proposed community detection model is t
X
introduced in detail. Section III introduces the model scheme Nic (t) = Ñic (τ ), (5)
and visual description of the proposed model. Section IV intro- τ =1
duces the experimental results, which verifies the advantages t 
X 
c c
of the proposed model for information propagation. Section V Eij (t) = Eji (t) = Ñic (τ ) + Ñjc (τ ) . (6)
summarizes this paper. τ =1

II. S YSTEM M ODEL Hence, community detection can be achieved by analyzing

the network domination matrix which records the process
This section introduces the details of the proposed com- of particle movements. Let δNc
(t) be the set of nodes of
munity detection model. The proposed model can complete community c at time t and we have:
semi-supervised community detection after several particle
c
movements. The element value of the network domination δN (t) = {vi | arg max (Nim (t)) = c, i ∈ V, m ∈ C}. (7)
m
matrix can be changed by the walking, splitting and jumping
operations of the particle. And the elements in the domination The proposed community detection model sets up the
matrix in turn affect the operation of the particles. operation of particle splitting ordering to solve the problem
in the previous particle competition aided models that the
A. Model Initialization number of particles belonging to different communities needs
In the initialization phase, giving a network represented by to be specified manually. Particles of different communities
G = (V, E), wherein V is the set of vertices and E ⊆ V ∗V is are allowed to determine the number of particles of that
the set of edges. Aij is the network adjacency matrix and aij community automatically based on the network topology by
is the weight of the edge between the node vi and vj . In an the splitting operation. The particles in the community c may
undirected graph, aij = aji . Suppose that the matrix S with split as they move. The adjacent node of the location of the
C ∗ V be the marked node set. c is a community number and particle pkc is denoted as vi and the judgment equation for the
Sc denotes the set of marked nodes of community c. The set split in vi can be denoted as:
of communities in the network is C = {1, 2, 3...c}, c ∈ C . In 
C C
each community, at least one node is labeled and one of the Nim (t) = 0,
P P
1, |Pm | < |V |,

nodes is selected as the labeled source node. Let L(c) be the ρi (t) = m=1 m=1 (8)
0, otherwise,

source node position of community c. Then, we can get
where |V | denotes the number of nodes in the network and
L(c) = arg max (vi |vi ∈ Sc : deg(vi )) , (1)
|Pm | is the number of particles in set Pm . If ρi (t) = 1,
where deg(vi ) denoting the degree of vi . In the initialization the splitting operation may be executed, and then, particle
phase, each source node can generate a particle. The total pk+1
c can be added to set Pc . In addition, according to the
model, the number of particles cannot be greater than that
of nodes in the network, so as to prevent too many particles
from being obtained by the operation of splitting. In addition,
in order to prevent a large number of different communities
of particles from crossing each other, the jump operation has
been set in the model. LP (pkc (t)) is recorded as the position
of the particle at time t. If the particle intends to move to vj
and LP (pkc (t)) = vi , the possibility of triggering the jump
operation can be formulated as:
!
c
c
Eij (t)
Pjump ij (t) = λ · 1 − P l
, (9)
l∈C Eij (t)

where the more particles in community c passed edge ij pre-

viously, the more possibility that other particles in community
c walking pass the edge and the less possibility of jumping
operation. λ ∈ [0, 1] is an adjustment parameter. Larger λ
makes it easier for particles to perform a jumping operation.
As λ approaches to 1, particles are more likely to jump, and
consequently, to stay out of the competition. If λ is 0, the
particles can not execute the jump operation and can always
randomly walk with full competition.
If the particle performs a jump operation, it can jump
back to the source node of the community which the particle Fig. 1. Example of particles walking in the model.
belongs to. If no jumps performed, it would move to node vj
successful. Therefore, the location of pkc at time point t + 1
can be given by: and community detection effectiveness into consideration, the
 model can stop particles’ movements and output the result
L(c), if jumps, when no nodes keep their domination vectors in the initial
LP (pkc (t + 1)) = (10)
vj , otherwise. state. At the end of each evolution, following equation should
After a jump operation performed, the particle is supposed be calculated:
to jump back to it’s community source node at the next time C
X
point. If no jumps performed, it would succeed to move to (Nim (t) − Nim (0)) 6= 0, ∀vi ∈ V. (12)
node vj . The community detection model adopts the method m=1
of dynamic source node adjustment ordering to adapt to the
dynamic evolution of the network. L(c)new is denoted as the If equation (10) is satisfied, particles can stop walking.
location of the new source node. Since the importance of nodes
in the network topology can determine the element values III. M ODEL A NALYSIS
of the domination vector, new core nodes can be determined
according to the domination vectors. The relationship between A. Particle Walking Example
L(c)new and the current source node location L(c) can be
In this section, the particle walking mechanism of the
formulated as:
proposed model is described intuitively. And Fig. 1 is an
L(c)new = arg max(vi |vi ∈ V : Nic (t)). (11) example of the model.
vi
As shown in Fig. 1, at time T, there are two communities
Every time the network evolves, the community detection of source nodes such as C1 and C2, three particles such as
model may retain the previous domination matrix and update P11 , P21 and P12 . There are seven nodes named V1-V7 whose
source nodes of different communities. Since the initialization domination vectors have been marked. When P11 moves from
is no longer performed, the particles in the model can continue V5 to V4, it needs to execute a jump according to the definition.
to walk and complete the community detection of the evolved So, the particle failed to move to V4 and returns to the source
network. Therefore, dynamic community detection for infor- node C1 at time T+1. P21 walks from node V6 to V5 and
mation propagation can be realized by the proposed model. reaches V5 at the time T+1, then, the node domination vector
of V5 has been changed. According to the definition, particle
C. Walking Completion P12 located at V2 is found that the domination vector elements
After several times of particles’ movements, the results of of the adjacent node V7 are all 0, so it executes a split
the community detection can be obtained. Hence, the stop con- operation. Then, the particle completes the split at T+1 and
dition needs determining. In order to take the time complexity changes the network domination matrix element values.
B. Introduction to Mathematical Analysis And Model Steps Algorithm 1 Community Detection Scheme
In this section, we are going to estimate the number of Input:
particles passing through the edge ij in different situations. A network composed of nodes and edges G, G = (V, E);
Firstly, the possibility of the number of particles passing Output:
through edge ij at time t is Ppass cij (t). In order to make each Community detection result and domination matrix;
1: Enter the nodes labeled by each community into the
particle own the same priority, the domination vector element
can be changed after all particles have moved. According to network;
2: Find the source node set C in each community;
the binomial distribution and the definition of particle jump
3: Perform particles walking:
probability, if there are n particles that intend to pass through
4: repeat
the edge ij at time t, the expected value of the number of
particles passing through edge ij at time t is set as Epass cij 5: Update the domination vector Ni (t), Eij (t) of each
which can be formulated as: node;
6: Determine whether the particle would “jump”. If yes,
Epass cij = n · Ppass cij (t) perform step 7, otherwise, perform step 8;
= n 1 − Pjump cij (t)


7: Make the particle jump to the core node;
c (13)
Eij (t) 8: Make the particle walk to the next node and determine
= n(1 − λ) + nλ P l
. whether to perform the split operation;
l∈C Eij (t)
9: until The conditions in equation (12) for stopping the
If one particle enters the edge at each time point, Ppass cij (t) walking are met;
would vary with time. Then, the average value of the prob- 10: return The domination matrix N (t) and community
ability of jump is set as Pjump cij (t). The expected value of detection result.
the number of particles passing through edge ij until time t
has been set as Epass cij (t). When t → ∞, supposing that ij is
no longer disturbed by other particles, the expected value of the network, the time complexity of that part is a constant,
passing particles in community c can be given by: in the initial phase. The state of the network domination
matrix is monitored in the process of particle walking. Because
lim Epass cij (t) = t · (1 − Pjump cij (t))
t→∞
!! the condition for stopping has been determined, the times
c of walking is a constant. The fifth step updates domination
Eij (t)
=t 1−λ 1− P vector of nodes and edges, the time complexity of this step
l c
l∈C,l6=c Eij (t) + Eij (t)
(14)
c
is O(|V | + |E|). In the process of particle walking, the
tλEij (t) number of particles produced by each type of source node
= t(1 − λ) + P l (t) + E c (t)
,
E
l∈C,l6=c ij ij is affected by the total number of nodes, so the total number
P l
of particles is proportional to |C||V |. |C| is the total number
where l∈C,l6=c Eij (t)=∆ , since it is not affected by other of communities. In the sixth step, when particles operate
particles, the constant ∆ can replace this term. Hence, walking, the average number of connected edges of each node
lim Epass cij (t) is 2|E|/|V |. Determining whether to perform a jump requires
t→∞  −1
(15) to check the values of the elements in the domination vector, so
∆
= t(1 − λ) + tλ +1 . the time complexity of this step is O(2|C||E|). The tenth step
(
c (0)+t 1−P
Eij jump ij (t))
c
is to perform |V ||C| operations when checking the domination
c vector and outputting the result. In conclusion, the running
 find that (1 − Pjump ij (t))
Then, we can  > 0 when t → ∞.
∆
time of the community detection model from data input to
Because lim +1 → 1, the expected result output is recorded as O(|V |+|E|+2|C||E|+|C||V |) ≈
t→∞ Eij (
c (0)+t 1−P
jump ij (t))
c

value of the number of particles passing through edge ij at O(|C|(|V | + |E|)). The time complexity of the model has
time t can be calculated as lim Epass cij (t) · t−1 → 1. a linear relationship with the number of nodes or edges.
t→∞
To sum up, assuming that particles from a community visit Therefore, it has good expansibility when applied to big data
an edge frequently. If the particle of this community intends to operations.
pass the edge again and t → ∞, the probability of the particle
IV. E XPERIMENTAL R ESULTS
successfully passing would approach 1.
Furthermore, in order to have a clear description to the In this section, the following five community detection
particle competition aided community detection model for methods for information propagation have been considered for
information propagation, the scheme of the proposed method comparison: (1)“Particle competition aided model”, which is
is described by Algorithm 1. proposed in this paper. (2)“LPA”, which is a classic dynamic
network community detection method [17]. (3) “ALPA”,
C. Time Complexity Analysis which is a new variant of the LPA [5]. (4)“DyPerm”, which
In this section, the time complexity of the proposed model is an algorithm that can solve the problem of dynamic
is introduced as a supplement. Because there are few nodes in community detection [6]. (5)“DynaMo”, which continuously
optimizes the modularity to obtain community detection re-
sults [7]. Furthermore, in order to describe the process of
information propagation clearly and verify the effectiveness of
the model proposed in this paper for information propagation,
three information propagation networks have been built. In
the networks, the “mentioned” behavior is adopted as the
connection between nodes.
The information propagation networks1 are as follows:
• The information propagation network about the terrorist
attack in Manchester (TAM network). The data in the
TAM network is from May 23 to May 31, 2017, on
the Twitter platform about that event. A total of 412,144
tweets about the discussion of the topic of the terrorist
attack in Manchester have been collected. Based on that
information, TAM network with 23,236 edges has been
built. Fig. 2. Comparison of the modularity of different dynamic community
• The information propagation network about the explosion detection methods on the TAM network.
at Parson Green subway station (EPG network). The data
in the EPG network is from September 27 to October
6, 2017, on the Twitter platform about that event. A
total of 25,696 tweets about the discussion of the topic
of Parson Green subway station explosion have been
collected. Based on that information, EPG network with
4,200 edges has been built.
• The information propagation network about the COVID-
19 which outbroke recently (COVID-19 network). The
data of COVID -19 network is the discussion about that
on Twitter on April 20, 2020. A total of 62,622 tweets
about the discussion of the topic of COVID-19 have been
collected. Based on that information, COVID-19 network
with 59,688 edges has been built.
In the comparative experiment, all edges are divided into
500 increments in chronological order. After that, the incre-
ments are continuously filled into the network. This paper
has evaluated the effectiveness of community detection results Fig. 3. Comparison of the modularity of different dynamic community
by modularity and evaluated the temporal smoothness [18] detection methods on the EPG network.
of community detection by Jaccard and NMI values. Since
the results are gradually flattening, the community detection
results of the first 50 increments in the information propagation
networks have been shown in Fig. 2, Fig. 3 and Fig. 4.
As shown in Fig. 2, the abscissa is the number of increments
during the evolution of the information propagation network.
The ordinate is the modularity of the community detection
result after each evolution of the network. Different lines
represent the modularity of the results of different community
detection methods.
The meanings of the elements in Fig. 3 and Fig. 4 are the
same as that in Fig. 2. Based on the results of the above
three networks, the results of particle competition aided model
are generally larger than those of other methods in terms of
modularity. In addition, modularity is one of the most reliable
metrics for measuring community detection results. Hence,
the particle competition aided model proposed in this paper
can solve the problem of dynamic community detection for Fig. 4. Comparison of the modularity of different dynamic community
detection methods on the COVID-19 network.
1 https://github.com/youguqiaomu/DPP
 TABLE I ACKNOWLEDGMENT
T EMPORAL S MOOTHNESS AVERAGE R ESULTS
The authors sincerely acknowledge the reviewers for their
Methods Temporal Smoothness Average Results suggestions which helped in improving the quality of the
Jaccard Values NMI Values paper. This research was funded by Competitive Selection
TAM EPG COVID-19 TAM EPG COVID-19
PCA 0.9887 0.9845 0.9903 0.9605 0.9613 0.9646 Project of Technical Research Programs of The Ministry of
LPA 0.9887 0.9118 0.9908 0.9529 0.9321 0.9576 Public Security (Grant No.2019JZX009) and People’s Public
ALPA 0.9915 0.9828 0.9924 0.9734 0.9723 0.9758 Security University of China “Special Project on Scientific
DyPerm 0.9885 0.9801 0.9907 0.9672 0.9572 0.9907
DynaMo 0.9908 0.9871 0.9918 0.9679 0.9647 0.9647
Research and Technical Innovation of Public Safety Behavior”.
Dr. Wang would like to acknowledge the financial support of
the Shuimu Tsinghua Scholar Program.
information propagation, which is better than other dynamic R EFERENCES
community detection methods. [1] D. Zhang, Y. Han, X. Li, “Dynamic Detection Method of Micro-blog
In the dynamic network, temporal smoothness is also an im- Topic Based on Time Series,” International Conference of Pioneering
portant metric for the quality of dynamic community detection Computer Scientists, 2018, pp. 192–200.
[2] X. Li, B. Wu, Q. Guo, X. Zeng and C. Shi, “Dynamic Community
results. In order to verify that the model proposed in this paper Detection Algorithm Based on Incremental Identification,” 2015 IEEE
not only improves the quality of dynamic community detection Int. Conf. Data Min. Workshop, ICDMW, Atlantic City, NJ, 2015, pp.
but also ensures that the value of temporal smoothness is 900-907.
[3] F. Meng, F. Zhang F, M. Zhu and Y. Xing, “Incremental density-based
within a reasonable range, we have calculated the temporal link clustering algorithm for community detection in dynamic networks,”
smoothness values of the above different community detection Math. Probl. Eng., 2016.
algorithms. Therefore, the average results of Jaccard and NMI [4] M. Zhu, F. Meng, Y. Zhou, “Density-based link clustering algorithm for
overlapping community detection,” journal of computer research and
values in the experiments of the three networks have been development, 2013, 50(12):2520-2530.
recorded and shown in Table I. [5] J. Han, W. Li, L. Zhao, et al., “Community detection in dynamic
In Table I, the particle competition aided community detec- networks via adaptive label propagation,” PloS one, 2017, 12(11).
[6] P. Agarwal, R. Verma, A. Agarwal, et al., “DyPerm: Maximizing
tion model proposed in this paper is abbreviated as “PCA”. permanence for dynamic community detection,” Pacific-Asia Conference
According to the results in Table I, when the proposed model on Knowledge Discovery and Data Mining. Springer, Cham, 2018, pp.
in this paper is adopted to complete community detection for 437–449.
[7] D. Zhuang, M. Chang, M. Li, “DynaMo: Dynamic Community Detec-
information propagation, the results of temporal smoothness tion by Incrementally Maximizing Modularity,” IEEE Transactions on
are similar to those of other dynamic community detection Knowledge and Data Engineering, 2019.
methods. In short, the proposed model also well-performed in [8] W. Li, Y. Gu, “Improvement of Stochastic Competitive Learning for
Social Network,” Computers, Materials & Continua, vol. 63, no. 2. 2020.
terms of temporal smoothness for information propagation. [9] J. Wang, C. Jiang, H. Zhang, et al. “Thirty years of machine learning:
The road to pareto-optimal wireless networks,” IEEE Communications
Surveys & Tutorials, 2020.
V. C ONCLUSIONS [10] J. Wang, C. Jiang. “Machine Learning Paradigms in Wireless Network
Association,” Encyclopedia of Wireless Networks, 2018.
In this paper, a community detection model has been pro- [11] M. Quiles, L. Zhao, R. Alonso and R. Romero, “Particle competition
for complex network community detection,” Chaos: An Interdisciplinary
posed. With imitating the phenomenon of information propa- Journal of Nonlinear Science, 2008, 18(3): 033107.
gation in the real world and combining the idea of emergence [12] T. Silva, L. Zhao, “Stochastic competitive learning in complex net-
theory, the model builds a particle competition aided scheme works,” IEEE Transactions on Neural Networks and Learning Systems,
2012, 23(3): 385-398.
for solving the dynamic community detection for information [13] F. A. N. Verri, P. R. Urio and L. Zhao, “Network Unfolding Map
propagation. Particles in the network may perform walking, by Vertex-Edge Dynamics Modeling,” IEEE Transactions on Neural
splitting and jumping according to the domination matrix and Networks and Learning Systems, vol. 29, no. 2, pp. 405-418, Feb. 2018.
[14] J. Wang J, C. Jiang, Z. Han, T. Q. S. Quek, R Yong, “Private information
these actions can also affect the network domination matrix. diffusion control in cyber physical systems: A game theory perspective,”
After several walks, the community detection results can be 2017 26th International Conference on Computer Communication and
obtained. Combining the time complexity analysis and the Networks. IEEE, 2017: 1-10.
[15] J. Wang, C. Jiang, T. Q. S. Quek, R Yong, “The value strength aided
simulation results, we can find that the model proposed in information diffusion in online social networks,” 2016 IEEE Global
this paper is more effective than other dynamic community Conference on Signal and Information Processing. IEEE, 2016, pp. 470–
detection methods in the community detection for information 474.
[16] J. Wang, C. Jiang, T. Q. S. Quek, X. Wang and Y. Ren, “The
propagation. Furthermore, the time complexity of the proposed Value Strength Aided Information Diffusion in Socially-Aware Mobile
model is linearly related to the number of edges and nodes, Networks,” IEEE Access, vol. 4, pp. 3907-3919, 2016.
so the model has good expansibility when applied to big data [17] N. Raghavan, R. Albert, S. Kumara, “Near linear time algorithm to
detect community structures in large-scale networks,” Physical review
operations. E, 2007, 76(3): 036106.
Our future work will employ the proposed model to develop [18] Y. Chi, X. Song, D. Zhou, et al., “Evolutionary spectral clustering
a more effective particle competition aided scheme that will by incorporating temporal smoothness,” Proceedings of the 13th ACM
SIGKDD international conference on Knowledge discovery and data
be used to complete the classification of nodes and graphs for mining. 2007, pp. 153–162.
static networks, multidimensional networks and heterogeneous
networks.

View publication stats