A method based on compressive sensing to detect
community structure using deep belief network
Liangliang Zhang, Haijia Wu, Jing Feng, Xiongwei Zhang
PLA University of Science and Technology
Nanjing, China
vermouthlove@hotmail.com, wu_haijia@163.com
Abstract—A deep learning scheme based on compressive
sensing to detect community structure of large-scale social
network is presented. Our contributions in this work are as
follows: First, we reduced the high-dimensional feature of social
media data via compressive sensing by using random
measurement matrix; Second, deep belief network is employed to
learn unsupervised from the low-dimensional samples; Finally
the model is fine-tuned by supervised learning from a small scale
sample sets with class labels. The effectiveness of the proposed
scheme is confirmed by the experiment results.
Keywords—compressive sensing; community structure; social
network; deep belief network(DBN)
I. Introduction
Over the past decade, research on community detection
has mostly focused on small-scale social network. These
algorithms can approximately divide into three categories: The
first one is based on graph theory including GN [1] and
FastGN[2]. Some other methods have been developed based
on matrix factorization. For example, symmetric nonnegative
matrix (SymNMF) has promising performance [3]. The
methods based on optimization called N-Cut and A-Cut were
proposed in [4].These methods can help us understand the
network structure and reveal the network functions and have
high efficiency when the size of network is small. However,
large-scale and high-dimensional network brings low
efficiency and the complexity of these algorithms grows
exponentially.
Compressive sensing (CS) [5][6] has outstanding
performance in processing sparse data. This theory has proven
to break the Nyquist sampling theorem in data sampling
process and is able to reconstruct the initial signal accurately
from the fewer projections by using the sparse priors of it.
Since the characteristic of social network is high-dimension
and large scale, the connections between nodes is relatively
less. For example, the sparsity of MovieLens dataset is 4.5%,
the sparsity of Netflix is 1.2%, and the sparsity of Delicious is
0.046%. Taobao has eight hundred millions commodities at
present, however, the number of products that an average
Taobao user browsed is less than 1000, so its sparsity is below
one millionth. To the best of our knowledge, the scale of
social network is larger, the dataset is sparser. CS is capable of
reducing dimension of dataset to extract the essential
information by utilizing random measurement matrix, so it is
suitable to handle high-dimensional network.
Deep learning (DL) [7][8] is a rising algorithm of
multilayer neural network. This method solves the problem of
local minimum and unsupervised training. DL adopts
unsupervised learning algorithm in training phase, and utilize
small labeled samples to supervised fine-tune the model in the
final phase. The representative set is selected from large-scale
social network, and its community labels can be obtained
through traditional algorithms mentioned above.
In order to learn community structure of large-scale and
high-dimensional network, we present a deep learning scheme
to detect community structure based on compressive sensing
in this paper. The proposed approach can improve the
efficiency of extracting community structure process from two
aspects. On the one hand, by using random measurement of
compressive sensing theory, we can reduce the dimension of
social network feature. On the other hand, by using deep belief
network (DBN)[9] model, the learning and prediction problem
of large-scale community structure become feasible.
II. Four Steps of Proposed Scheme
The overall procedure of the proposed community
detection scheme is shown in Fig. 1. The scheme has four
steps: Firstly, community structure of the representative set
which extracted from original adjacency matrix is detected via
GN algorithm. Then, the dimension of original adjacency
matrix is reduced via random measurement matrix.
Unsupervised training of DBN is implemented by utilizing
low-dimensional feature samples. Finally, the small set of
labeled samples that obtained in the first step is used to fine-
tune the DBN with supervision.
Adjacency
Matrix
Represent-
ative Set
Take
Samples
GN Algorithm
Random
Measurement
Compressed
Feature Matix
Labeled
Samples
Pretrain
DBN
W
1
W
2
V
H
1
H
2
DBN
W
c
H
2
Classifi-
cation
Layers
DBN
Supervised
Fine-tune
CD Algorithm
Take
Samples
BP Algorithm
Fig.1. Overall procedure of the algorithm
International Workshop on Cloud Computing and Information Security (CCIS 2013)
© 2013. The authors - Published by Atlantis Press