Research On Community Detection In Complex Networks Based On Local Feature Diffusion

With the rapid development of information technology, complex network analysis has become a hotspot in machine learning and data mining. Research shows that community structure can reveal some underlying characteristics and rules in complex networks. Therefore, community detection is of great significance for mining network structure information and understanding the characteristics of network structure.In the research of network community structure detection, how to describe the characteristic matrix of network has a great importance on community detection. At present, most researches are carried out from two aspects: the local features and the global features, and analyze the undirected graph with symmetric features. In this paper, we focus on an idea of combining the local features and global features to propose a new construction method of the network characteristic matrix. The main contents are as follows:(1) A new method of constructing characteristic matrix based on local feature diffusion is proposed. First we redefine node similarity, then enhance local features with weighted links, and finally spread these features by diffusion kernel to get the network characteristic matrix. The new characteristic matrix, which combines the local and global features of the network, is a more comprehensive representation of the topology information.(2) A relatively stable label propagation algorithm is proposed. Based on the characteristic matrix which is constructed by local feature diffusion, we first strengthen the local structure of the network with weighted links, then construct the label contribution matrix with the new characteristic matrix, and finally give priority to update the labels of the core nodes in the first update phase. This algorithm can not only guarantee the accuracy, but also enhance the stability of the algorithm.(3) The idea of local feature diffusion is applicable to non-negative matrix factorization model, and two new characteristic matrix factorization models are proposed. Firstly, the newly defined symmetric characteristic matrix is applied to symmetric non-negative matrix factorization model. Then aiming at the deficiency of the symmetric similarity measurement method describing network topology information, we propose a construction method of asymmetric characteristic matrix, and combine it with non-negative matrix factorization model for community detection. In this method, the local similarity of the nodes is measured asymmetrically, and the undirected graph is adjusted into a directed weighted graph. Then the local features of network are enhanced by weighted links. Finally, we spread these features by diffusion kernel to construct an asymmetric characteristic matrix. Experimental results show that the symmetric characteristic matrix and asymmetric characteristic matrix we proposed can effectively improve the effect of community detection in complex networks, and have a unique advantage in small-scale community identification resolution. Especially the asymmetric characteristic matrix, which can better measure the similarity between the nodes, improve the performance of the community detection obviously.