Research Of Community Detection Based On Non-negative Matrix Factorization

Community detection is an important part of complex network analysis.The community structure division of network through community detection algorithm can help people dig out the structural characteristics and related functional characteristics of complex network.In recent decades,scholars have proposed a large number of community detection algorithms,but these algorithms usually only focus on the topology information of the network,and can't effectively partition the outliers to suitable communities in the incomplete graph.By an in-depth analysis of existing community detection algorithms,this thesis proposes two new community detection algorithms based on the combination of node feature information and network hierarchy information,aiming to improve the accuracy of community detection.The main research contents are as follows:(1)In order to divide outliers in social networks into appropriate communities,this thesis proposes a community detection algorithm,CSNMF,which combines node feature information and topology information.CSNMF uses a self-encoding non-negative matrix factorization framework composed of decoders and encoders to integrate the topology information and node feature information of the network.It set an adaptive parameter based on the similarity of the topology information and node feature information to effectively control contribution rate of node feature information to community division.By doing so ensures that the node feature information effectively supplements the topology information,thereby improving the accuracy of community detection.In addition,CSNMF retains the good interpretability of the non-negative matrix factorization framework,and can accurately describe the characteristics of the divided community structure.(2)In addition,this thesis proposes a deep autoencoder-like non-negative matrix factorization model(DSCNMF)that combines node feature information and topological structure information to further improve the accuracy of community detection.Through the deep autoencoder-like framework,DSCNMF makes a hierarchical analysis of the topology information of the original network,so as to learn the hierarchical mapping between the original network and the final community division.It can also learn the hidden hierarchical information of the original network from low to high in the middle layer.DCSNMF not only utilizes the node feature information,but also models the node feature information into the deep encoder-like framework.At the same time,the contribution rate of node feature information is controlled by adjusting coefficient,so as to ensure community detection performance of the algorithm for incomplete graphs.In order to verify the effectiveness of CSNMF and DCSNMF,this thesis conducts experiments on several real social network data sets.Compared with the experimental results of latest related community detection algorithms,both CSNMF and DCSNMF can effectively mine the community structure in the network,and DCSNMF has better community detection performance in the network with rich hierarchical information.