Research Of A Graph Neural Networks Based Community Detection Algorithm For Complex Networks

Community is a kind of structure with internal close connection and external sparse connection in complex networks.Community detection is to find communities in complex networks.Graph neural network is a kind of neural network used to process graph data.Graph convolutional networks(GCNs)and graph attention networks(GATs)are gradually used in community detection tasks in recent years.In addition,there is a derivative method which integrates the Markov random field(MRF)into the convolution layer.However,they can only extract the first-order neighborhood information of nodes in the same layer,and can not take into account the impact of high-order neighborhood of nodes.There is also a problem that the design of energy function is not comprehensive in the derivation method of MRF with convolution layer.Therefore,an end-to-end community discovery method based on GCN,GAT and MRF model is proposed.This paper designs a multi-level graph attention mechanism,through which the higher-order neighborhood information of nodes can be extracted in the same network layer,which solves the drawback that the current graph attention network can only pay attention to the first-order neighborhood information of nodes in the same layer,and can avoid the transition smooth problem caused by deepening the network.An aggregation strategy is designed to aggregate the extracted multi-level neighborhood features as the output of the layer.An improved MRF model is introduced into the last layer of neural network,and an expression formula of energy function is designed.An improved joint similarity calculation method is proposed for the similarity calculation between nodes in pairwise potential.This method combines topological structure similarity and node feature similarity,and considers the influence of node influence on community division.The community rough partition is refined by minimizing the energy function.Compared with two mainstream algorithms of graph neural networks: GCN and GAT as well as a method combining graph convolution networks with MRF(MRFas GCN),five artificial network data sets(LFR0-LFR4),seven real network public data sets(including three paper reference networks and four university Web page reference networks)and a city network data set are tested,and multiple groups of community division results are obtained.The experimental results show that the proposed method has better performance for accuracy,standardized mutual information(NMI)and other community division evaluation indicators.