Study On Unsupervised And Semi-supervised Stochastic Blockmodel

Many complex systems in nature and human society can be abstracted to complex networks in nature and human society in order to gain valuable results.Therefore,complex networks are widely used in various fields,such as biology,social relationship and Internet and so on.Depending on the type of edges,complex networks can be divided into unsigned networks and signed networks.In unsigned networks,there are two types of links: positive and nonexistent links;in signed networks,links fall into positive,negative and nonexistent links.As an important statistic model in complex networks,stochastic blockmodel has gained a lot of attention of researchers because of its flexibility.However,in real-world,noise or the sparse problems are existed in real-world networks,so there are a lot of challenges in the research of stochastic blockmodel for complex networks.To solve the above problems,this paper makes the following contributions:1)A sign prediction method is proposed based on the signed stochastic blockmodel and,the model is completely validated in community detection and sign prediction.Based on Bayesian theory,the posterior distributions of the three type links are deduced by using the prior distribution of the edges and the results of community detection and then use the posterior distributions to infer the sign of an unseen edge.This method performs better than other sign prediction method because it takes into account both prior distribution and the link density and sign.Besides,the performance of the model is completely tested,including the ability of community detection with synthetic and different scale real-world networks,the model selection capacity with both balanced and unbalanced networks,and the ability of sign prediction with synthetic and real-world networks.2)A general semi-supervised stochastic block model is proposed.This model can detect community and multipartite structures with both signed and unsigned networks.Since there are noise in networks and networks are sparse in real-world,detecting hidden structures by unsupervised methods is very difficult.In many real scenarios,node labels can be partially gained.By simultaneously using label and topology information,a novel general semi-supervised stochastic block model and variational Bayesian inference learning method are presented.Utilizing the statistical nature of SBM,this model also detect both the community and multipartite structures.It is the first semi-supervised method for mining structures in signed networks.The model is tested by comparing with unsupervised,semi-supervised and supervised methods for detecting community and multipartite structures.