Research On Community Detection Based On Deep Learning

Social network is a network that contains a large number of nodes and intricate relationships among nodes,and there is a clear community structure among social net-work.The community refers to the sub graph gathered by nodes in the network.The nodes in the same community are closely connected while the connections among nodes in different communities are sparse.Researches have been conducted for community detection and related algorithms have been proposed.So far the algorithms,such as random model or the modularity maximization method,essentially map the network structure to a feature space for representing and reconstructing the topology of the network.However this mapping is linear,and the real-world network contains a large number of nonlinear features,which makes the above models not effective in practice.In the community,there are some nodes that are not closely connected to the main part of the community.These nodes may be single nodes which will be community outliers,or may be a small group formed by several connected nodes,that is,the subcommunity-outlier.Subcommunity-outliers may represent some meaningful information,but there are few related researches on subcommunity-outlier.We focus on the above questions,and put forward the improvements.The main contributions are as follows:Firstly,aiming at overcoming the shortage of conventional community detection algorithm,a community detection algorithm based on deep learning is proposed in this paper.This method combines the advantages of Convolutional Neural Network and Auto-Encoder,and constructs a deep learning network ConvAE.This model can extract the nonlinear characteristics of the network and the local perception of convolution op-eration is also consistent with the local similarity in the network,thereby improving the effect of community detection.To better characterize relations between nodes,we propose similarity and employ similarity matrix as the input of ConvAE.In the ex-periments,we compare our method on synthetic and real-world networks with tradi-tional methods and deep learning algorithms.On the majority of datasets,our method ConvAE outperforms other algorithms.Meanwhile,we compare our representation on different clustering algorithms to verify our network's ability of representation.Secondly,to address the problem about description and definition of existing com-munity detection do not well reflect outliers,this paper proposes a subcommunity-outlier detection method which based on deep learning.The subcommunity-outlier is a subcommunity which connects with the main part of community sparsely.We re-define the edge betweenness as the degree of closeness between two subcommunities in the community.A network combining convolution operation and Auto-Enoder is constructed to get the value that can represent the network.In the experimental section,we propose three evaluation functions to evaluate our methods and two baselines for comparison.We verify our method on synthetic datasets and real-world datasets.Fi-nally,the subcommunity-outliers in real-world network have been analyzed detailedly to indicate the significance of subcommunity-outlier detection.Thirdly,with the development of the Internet,the network structure is becoming more and more complex.In order to deal with the complex and huge network more effectively,a community detection algorithm based on parallel deep learning process is proposed in this paper.The community detection algorithm proposed in the first part is processed in a distributed TensorFlow framework,and the training process of deep learning is distributed to multiple nodes by setting the parameter server and work node.Thus,it can improve the efficiency of the algorithm and reduce the running time.In the experiment part,we compare the time of the algorithm in different data sets and different number of worker.At the same time,we also compare the results of community detection in single and distributed cases.All the experiments verify that the algorithm reduces the running time while ensuring the community detection effect.