Research On Community Detection Based On Combination Of Network Multi-topology Characteristics And Attributes

Community detection has been a very popular research direction in the field of data mining,and is widely used in social network analysis.However,most of existing researches are directly on single layer network for cutting or hierarchical clustering,ignoring the existence of a variety of structural characteristics of social networks.With the development of social network,the connection relationship between users is heterogeneous,showing a multi-layer structural characteristic;the influence of different users is not the same,and there are differences in the degree of importance,which is characterized by the unbalanced weakly bipartite structure of community kernel and auxiliary community.At the same time,according to clustering effect of social network,the common neighbor attribute can be used as an important basis for community detection.Aiming at the problems of existing community detection approaches,a community kernel based random walk community detection framework is proposed.It effectively uses characteristics of multi-layer structure,community kernel and auxiliary community structure and common neighbor attribute of social networks,is an approach for community detection containing two stages,including detecting community kernels and finding their respective auxiliary communities.In order to detect high quality community kernels,a novel measurement MACM based on multi-layer structure and attribute information is proposed,which includes node attribute similarity measureS_A(u,v)and node importance measure I(u).Then,two improved community kernel detection algorithms based on MACM are proposed,including MA-Greedy which is based on greedy idea and MA-WeBA which is a weight balanced algorithm.The multi-layer structure and attribute information are comprehensively considered when calculating similarity and importance of nodes.Results show that multi-layer structure and attribute information are able to enhance precision of community detection,and the algorithms are scalable.In order to make extended communities closer to real situation,an initial community expansion method based on random walk is proposed,which transforms the extension of initial community to a classification problem.By constructing random walk graph,then basing on random walk probability model to divide nodes to initial community kernels,so as to improve the performance of community detection.At the same time,the approach can detect communities with better uniformity.