Research On GCN With Neighborhood Selection Strategy

Convolutional neural networks(CNN)have achieved great success in image representation and processing.In recent years,they have also received a lot of attention in the field of graph mining,and research on graph convolution networks(GCN)has emerged.Since graph often does not have an ordered data structure like image,the neighborhood of the nodes to be encoded has a crucial role in the spatial domain-based GCN model.Considering the different structural influences of node neighborhoods,this paper proposes a graph convolutional network model based on neighborhood selection strategy,named Co N-GCN.The model first collects structurally important nodes in neighborhood for each central node and performs hierarchical selection to obtain the core neighborhood.Then the features of the central node and its core neighborhood are sorted by the structure importance to obtain a matrix,and finally input to the deep CNN for semi-supervised learning.Experimental results show that the performance of Co NGCN is superior than GCN and LGCN in node classification task with real network datasets Cora,Citeseer,Pubmed.Considering that the network evolves dynamically,this paper also proposes a dynamic version of Co N-GCN,that is,a time-series graph convolutional network model(TS-GCN).The model first collects the edge timestamp in a fixed time interval,sorts the neighbor nodes in chronological order and obtains the core neighborhood,and finally forms an input matrix of a deep convolutional network for semi-supervised learning.The experimental results show that the TS-GCN model performs better than the static GCN model in both node classification and link prediction tasks with Epinions,DBLP,Academic datasets.In order to extract deeper features of the nodes and extend the current algorithm to large-scale networks,this paper finally proposes corresponding deep models and subgraph training methods.The experiments prove that with the deepening of the network,the accuracy of the method has been further improved in the node classification task;on the large graph training task,the improved method can effectively reduce the training time while ensuring the accuracy.