Research On Community Detection Based On Variational Bayesian Inference

Community structure is one of the most important characteristics of complex networks,usually in the form of “nodes connect tightly within the community,and connect sparsely between communities”.Bayesian model is a mainstream community detection method.Its basic idea is to establish a probabilistic generative model to fit the real network,and infer the best parameters.However,existing models and their inference algorithms generally have problems such as low efficiency and overfitting,and are not accurate enough for the cognition and analysis of the communities.This article studies how to build more complete Bayesian models and inference algorithms,and conducts research from two aspects:(1)We extended the network embedding-enhanced Bayesian model for generalized community detection.Different from traditional methods which only focus on the assortative network structure,the model adopted the concept of generalized community,including disassortative and other mixture structures.At the same time,to deal with the noise and redundant information in the topology,the network embedding matrix was used to represent the nodes as feature vectors in low-dimensional space.Expanding on this model,we integrated the generation of model parameters into the overall generation process and introduced the corresponding conjugate prior distributions,to build a complete and unified hierarchical Bayesian model.We then designed the variational inference algorithm to estimate the parameters,effectively improving the accuracy and applicability of the model.(2)We extended the community detection model combining network structure and node attributes.Nodes in networks often carry rich attribute information,which can not only reveal node characteristics,but also improve the quality of community detection.To fully utilize node attributes,a joint model for topology community detection and attributed topic discovery can be constructed.A correlation matrix is used to improve the matching between them.We integrated model parameters related to community memberships into the modeling,then constructed an extended hierarchical Bayesian model based on conjugate priors.The variational inference algorithm was given to update parameters iteratively.Experiments showed the method can maintain high efficiency and accuracy on large-scale networks.It can also derive multi-semantics of communities,showing a good description ability.