Research On Community Detection Methods Based On Probabilistic Graphical Model

Probabilistic graphical model is a powerful tool for solving complex problems in the real world,which includes directed graphical models(i.e.Bayesian model)and undirected graphical models(i.e.Markov Random Field,MRF).In recent years,they have been widely used in the research of community detection.However,there are still two problems.1)Community detection methods based on directed graphical models often ignore the diversity of community patterns in the real world,and use network topology which is often noisy and sparse.2)As another major category,the research of MRF applied to community detection is still in its preliminary stage.In order to solve the above problems,we make a deep research about community detection based on probabilistic graphical model.First,we construct a network embedding-enhanced model for generalized community detection based on directed graphical model.This model uses an idea of mixture modeling to describe network regularities,and introduces network embeddings to further enhance its ability to describe network communities.So it can not only find generalized communities,but also be robust for the sparsity and noisy of networks.Second,we construct a MRF-based community detection model,which combines node attributes and network topology.The model obtains topical cluster distributions by LDA and redefines unary potentials based on it,then redefines pairwise potentials based on network topology.So this new MRF model can effectively combine node attributes and network topology.We also present effective model inference algorithms for the above two models to perform parameter optimization.The experimental results demonstrate the superior performance of our algorithms in the realm of probabilistic graphical models.In summary,we propose two different community detection models based on the directed as well as undirected graph models,respectively,which are included in the probabilistic graphical model,and design the suitable model inference algorithms to optimize the parameters.They can solve problems of existing community detection models and own good usability in complex real-world networks.