Research On Semi-supervised Node Classification Based On Propagation Completeness And Directionality Of Graph Convolutional Networks

Graph is a kind of spatial topological structure which is ubiquitous in real life.It has been widely used because of its special structure mode,but there are some difficulties in dealing with it.Graph convolutional network can deeply mine the node information and edge information in the graph,filling the gap in topological structure data processing,so it is favored by the majority of scholars,among which the node classification problem has important research significance.In the current study of graph convolutional networks,there are two problems: one is the balance between node information integrity and network training efficiency of graph convolutional networks.Most of the existing Graph Convolutional Networks,such as Graph Convolutional Networks(GCN)and Simplifying Graph Convolutional Networks(SGC),involve a small neighborhood range during training,resulting in information loss of higher-order neighbor nodes.However,Graph convolutional Networks that can completely retain Graph structure information,such as Graph Attention Networks(GAT),have slower training speed and consume more resources.Another problem is the application of spectrogram theory to digraph.Most of the existing graph convolutional networks relax the directed graph structure into an undirected graph so that spectrum decomposition can be correctly applied.This will lead to the destruction of the graph structure,resulting in the loss or redundancy of important information and affecting the learning task of the graph.In view of the above two problems,this paper has done the following research:(1)Aiming at the first problem,this paper designs a graph convolution network model based on feature accumulation.Firstly,the higher power of the adjacency matrix is defined according to the probability transfer function based on random walk,which is used as the feature propagation path of neighborhood nodes in different ranges.Then,the new Laplacian operator is used to carry out convolution operation in different neighborhoods.Finally,the obtained feature information of different levels is fused to achieve the purpose of preserving node features.Experimental results on real data sets show that the proposed model achieves high classification accuracy in a certain neighborhood and shorts the training time.For example,in Pubmed data set,the classification accuracy is improved by2.5% compared with the benchmark algorithm,and the training time is shortened from GAT ’s more than an hour to tens of seconds.(2)For the second question,this paper logically directed graph view for the view and the degrees of split as two separate figure structure,according to the defined with the degree of probability transfer matrix and the degree of Laplacian matrix,and then respectively to a convolution of two views,said new feature is obtained by integration function.The experimental data on three directed citation network data sets are compared with the benchmark algorithm.The results show that the proposed method can effectively improve the classification accuracy of directed graphs.