Community Detection And Role Discovery Based On Multi-layered Networks

The rapid development of information technology provides new methods for network research.As two important directions in the field of network research,community detection and node role discovery have gradually attracted researchers' attention in recent years and they are widely used in different fields.Networks in the real world are often multi-layer networks.Compared with single-layer networks,multi-layer networks can contain more information and reflect the real situation to a greater extent.This thesis focuses on multi-layer network community detection and the role discovery of directed weighted graphs.The main contributions include the following points:1)The existing public data sets on multi-layer network community detection and node role discovery tasks lack community and role labeling information.This article obtains the real relationship network between students and teachers in a university based on the actual situation,and uses Markov Monte Carlo(MCMC)algorithm generating communication data between them,building a fully labeled multi-layer network data set HUST.2)An auto-encoder community detection method based on multi-layer network is proposed.This method considers the importance of each layer and associates it with the similarity between the layers of the multi-layer network.By calculating the similarity between the layers of the node,the importance of each layer in the multi-layer network is obtained.Aiming at the problem that the node similarity method based on the independent path ignores the closeness of neighbor nodes,and the long independent path interferes with the calculation of similarity,a new node similarity calculation method is proposed,this method performs pruning and wighting on the path that is too long.The autoencoder is used to better extract the features of the multi-layer network,and the effective low-dimensional vector representation of each node of the multi-layer network is obtained.Finally,the low-dimensional vector of these nodes are clustered through the clustering method,and the result of the clustering is the final community division result.3)For undirected unweighted networks,a method for calculating the centrality of the eigenvectors of the similarity matrix is proposed to obtain the importance of undirected unweighted network nodes,and for directed weighted networks,a multi-index measurement method is proposed to measure the nodes on the directed weighted network.Finally,the two parts of the undirected graph and directed graph are merged,and the node roles of the multi-layer network are divided by the clustering method.