Research On Key Technologies Of Graph Neural Network

In recent years,deep learning has made great progress in computer vision,natural language processing and other fields.However,in the field of structured data,deep learning based methods do not perform well.Since graph data contains rich information,the analysis and processing of graph data has been regarded as an important research field.Graph neural network(GNN)is a recently proposed framework for modeling and learning graph data,which directly use deep learning technology.Its excellent performance has aroused great attention and in-depth exploration of researchers.The progress has been made in the field of GNN,but there are several problems that need to be solved.This paper mainly focuses on two key problems of graph neural network:(1)How to expand the depth of GNN and improve the performance of GNN by expanding the depth of GNN.The existing GNN are based on the message-passing mechanism,which results in performance degradation and indistinguishable nodes when several layers are stacked.Therefore,existing graph convolution methods only contain 2 or 3 layers.(2)How to solve the problem that the existing methods perform as low-pass filters on the graph and improve the performance on the graph with high heterogeneity.The existing GNN only retain the low-frequency information of the nodes in the graph by summing up the neighbor features,which leads to the problem that the node features become similar and perform poorly in the graph with high heterogeneity.Both of these problems hinder the development and application of graph neural networks.The graph classification methods with graph pooling layers have the problem of information loss in the process of down-sampling because of insufficient network depth.To tackle this problem,this paper proposes the residual convolution graph neural network(RCGNN).RCGNN improve the performance by expanding the network depth,a reasonable down-sampling strategy and a new extracting method of graph-level feature.Extensive experiments on five datasets containing a large number of graphs show that RCGNN outperforms all tested methods.Ablation experiments are also provided in this paper,which provide a basis for the interpretability of RCGNN.The graph convolution methods aggregate features of neighbors by summation of neighbor features.Those methods essentially perform as low-pass filters and fail to make full use of the high-frequency and low-frequency information of neighbors,which result in poor performance on heterophilic graphs.In this paper,the negative graph attention network(NGAT)is proposed to explore the role of high frequency information in updating nodes.features.Furthermore,this paper proposes the high-frequency and low-frequency graph attention network(HLGAT),which makes full use of the high-frequency and low-frequency parts of neighbor features through two channels.In this paper,Extensive experiments on five heterophilic graph datasets are conducted.The experimental results show that the HLGAT outperforms all tested methods.Ablation experiments are also provided in this paper,which provide a basis for the interpretability of HLGAT.At the end of the paper,the research is summarized and the future research is prospected.