Research For Community Detection Integrating Attribute And Link Relationship

Community detection has great significance of complex network analysis.Through analyzing the amount of implicit information in the networks,the hierarchical structure among members can be accurately detected.Community detection has captured a large amount of attentions nowadays,and rich theoretical technologies are the foundation of deep research on community detection.Based on existing research,the community detection methods that combine attributes and linked information are studied,and two methods of community detection in different application scenarios for a great deal of attribute information existed in complex networks are proposed.To improve the density and efficiency of community detection,the Bayesian and Non-negative Matrix Factorization are also employed.The main contents of this thesis are shown as follows:1.To improve the density of overlapping community in attributive complex networks,an improved community detection algorithm based on Unified Bayesian Nonnegative Matrix Factorization is proposed.First,to combine the attribute and link information effectively,the unified Bayesian network generation model is improved.Then,the Bayesian theory is introduced to analyze and verify the objective function.Last,the final iterative update rules is obtained according to the Non-negative Matrix Factorization.The experiment shows that the community detected by this algorithm yields a relatively high density and has a strong structure,which is suitable for the overlapping community detection in large communities in small networks.2.To improve the efficiency of non-overlapping community in attributive complex networks,an improved Community Detection algorithm based on Unified Bayesian Clustering is proposed.First,the unified Bayesian probability model based on LDA is presented,and an iterative update formula based on Bayesian probability model is obtained.Then,the improved random walk algorithm is proposed,which is used for initialization of the community.Last,the algorithm is obtained through combination of the update rules and the initialized community.The experiment with real datasets shows that the algorithm has a good performance in non-overlapping community detection and it is suitable for community analysis of the network that only contains linked information.The two algorithms proposed in this thesis yield good performance in overlapping and non-overlapping community detection respectively.Both of these two methods are integrated with the attribute and linked information in the networks analysis,which is of great significance to the community detection research that combines link and attribute information.