| Complex network is composed of nodes and edges,which is helpful to understand the network structure of reality.One of the most important directions for the study and understanding of complex networks is community discovery,through the study of communities,not only can we get the existing topology of the network,but also know the future evolution direction of the network;In recent years,the results of community discovery have also been widely used in the fields of recommendation systems and public opinion analysis.Community discovery can be divided into non-overlapping community discovery and overlapping community discovery.However,there are currently inefficient divisions of non-overlapping community discovery algorithms,fewer overlapping community algorithms,and fewer algorithms suitable for large-scale network community discovery.Graph neural networks(GNN)are introduced in this paper,because in the widespread use and development of deep learning,graph neural networks have shown good use and can achieve reliable performance on multiple graph downstream tasks such as node classification and link prediction.This paper uses the network embedding node vector representation,combined with GNN to achieve non-overlapping community discovery and overlapping community discovery.The main research content of this article includes the following four parts:(1)A representation learning method BC-Node2 vec for overlapping community discovery is proposed.According to the structural characteristics of overlapping nodes,this paper improves the wandering strategy of Node2 vec algorithm,introduces the concept of node intermediates,and performs biased overlapping node wandering to generate characteristic vectors of nodes.Experimental results show that compared with the traditional representation learning method,the algorithm has better representation ability for overlapping nodes.(2)A plot convolutional model based on agglomeration coefficient C-GCN is proposed.The agglomeration coefficient can reflect the density of the node neighbor connection,the node agglomeration coefficient is fused with the adjacency matrix,as the initial input of the graph convolutional network,in order to better react to the node information,the attribute matrix in the GCN is the modularity matrix,and the improvement of the graph convolutional neural network is realized.After the node classification comparison experiment with GCN,it is proved that C-GCN has better network characterization ability.(3)NC-GCN,a community discovery model based on local and global information extraction,is proposed.In this paper,according to the proposed Node2 vec algorithm,the local information of the network is obtained,the global information of the network is obtained according to the C-GCN,and the obtained information is fused to obtain the final network node representation,and finally the community discovery is achieved according to the k-means clustering.By comparing with the classic community discovery algorithm,it is confirmed that the model has better performance for community discovery tasks.(4)BC-GAT,an overlapping community discovery model based on graph attention network,is proposed.In this paper,the BC-node2 vec algorithm is used to obtain the eigenvectors of nodes in the network,and then the characteristics of the nodes obtained by BC-node2 vec as the initial node vectors of the graph attention network,and the multi-head attention mechanism is used in combination with the adjacency matrix to learn the community membership relationship matrix,and finally the maximum negative log likelihood estimation of the Bernoulli-Poisson(BP)model is used to learn the real community membership relationship matrix.In addition,nodes are assigned to communities according to the matrix values,and finally overlapping community discovery under high-dimensional data is realized.Comparative experiments have shown that this algorithm has a better effect on overlapping communities. |
