Abstract: With the development of social network, users in all kinds of social networks form a large user graph. The articles published by users and the links between articles form a large document graph. According to the user graph and document graph, how to mine the topics, communities and user influence is an important problem, but now this problem is handled independently. Considering the inter-dependence of the problem and making use of text content and coauthor relationship, this paper proposes a joint model of topic modeling, community discovery and user influence analysis based on LDA (latent Dirichlet allocation), called ACT-LDA (author-community-topic LDA), which uses variation method for inference. The experiments on DBLP data show very good results, which validates the proposed model.

Key words: social network, community discovery, topic modeling, user influence

摘要： 随着社交网络的发展，社交网络中的用户形成大规模的用户关系图，用户在社交网络中发表内容，这些内容及其链接关系形成大规模的文档图。如何根据用户关系图、文档图，挖掘出用户所形成的社区、社区用户的影响力以及各个社区的话题，是重要的问题，而目前这些工作相对独立。考虑了用户发表内容、用户之间的关系信息，利用话题传播、社区形成和用户影响力之间的关联性，提出了一个基于LDA（latent Dirichlet allocation）的集成话题发现、社区发现和用户影响力分析的统一模型ACT-LDA（author-community-topic LDA）。模型采用变分推理的方法解决推理问题。在DBLP数据上进行了实验，取得了非常好的结果，证明了模型的有效性。

关键词: 社交网络, 社区发现, 话题模型, 用户影响力