MEI: Mutual Enhanced Infinite Generative Model for Simultaneous Community and Topic Detection

Abstract

Community and topic are two widely studied patterns in social network analysis. However, most existing studies either utilize textual content to improve the community detection or use link structure to guide topic modeling. Recently, some studies take both the link emphasized community and text emphasized topic into account, but community and topic are modeled by using the same latent variable. However, community and topic are different from each other in practical aspects. Therefore, it is more reasonable to model the community and topic by using different variables. To discover community, topic and their relations simultaneously, a m utual e nhanced i nfinite generative model (MEI) is proposed. This model discriminates the community and topic from one another and relates them together via community-topic distributions. Community and topic can be detected simultaneously and can be enhanced mutually during learning process. To detect the appropriate number of communities and topics automatically, Hierarchical/Dirichlet Process Mixture model (H/DPM) is employed. Gibbs sampling based approach is adopted to learn the model parameters. Experiments are conducted on the co-author network extracted from DBLP where each author is associated with his/her published papers. Experimental results show that our proposed model outperforms several baseline models in terms of perplexity and link prediction performance.