Integrating Topic Models Into Markov Random Fields For Community Detection

Markov Random Field(MRF),which belongs to the undirected probability graphical model,is a tool that can effectively model network data in the real world.Recently,MRF has been successfully applied to community detection and has good results.However,most of the existing MRF-based community detection methods only use topology data in the network for community detection,while the attribute information which also plays an important role,is ignored.To solve this problem,this paper integrates the topic model that is good at describing attribute semantics into the MRF community detection model that is good at describing topology data.Firstly,this paper uses topic clustering model to fit the attribute information according to the two-stage approach,making it could use the topology information and attribute information at the same time.It enhance the ability of community detection of the model.However,this two-stage approach cannot realize parameter sharing between sub-models,so it cannot truly realize the complementary advantages of the two types of information sources.An ideal approach is to build a joint model by an end-to-end way to further enhance the ability of community detection.However,the topic models belong to directed probability graphical model while MRF is an undirected one.They have different properties and training approaches so it is very challenge to build the parameter sharing mechanism of the model.Aiming at this problem,this paper proposes a new method which could integrate topic models into MRF by an end-to-end way according to factor graph model.And then developed a set of belief propagation algorithm based on factor graph to optimize the model parameters.It makes the model truly realize the mutual adjustment of parameters between the two molecular models,and make full use of the advantages of the two models.It further improve the performance of community detection method based on MRF.