Graph Convolutional Network Based On Structure Enhancement

Graphs exist widely in various real-world applications,and with the rise and rapid development of web networks,social networks,biochemical molecules and other fields,research on graph-structured data has received widespread attention and directly driven it to become a popular area of academic research in recent years.Although traditional deep learning methods have been applied to extract features from Euclidean spatial data with great success,the performance of traditional deep learning methods on processing nonEuclidean spatial data is hardly satisfactory.In recent years,graph neural networks,especially graph convolutional networks,have received a lot of attention.Among them,graph structure determines the connection between nodes,which is the key in graph convolutional networks.This paper presents an in-depth study of graph convolutional networks,and the main findings of the paper are as follows:(1)To address the over-smoothing problem of graph convolutional networks,this paper proposes a new framework high-order graph attention network(HGRN)for structure enhancement to optimize the neighbor aggregation process.HGRN uses the attention mechanism to obtain the connections between nodes and multi-hop neighbors,adaptively learns the importance of nodes with different hop neighbors and performs neighbor aggregation.The model learns the implicit graph structure information to make full use of the high-order information in graphs.Meanwhile,HGRN performs similarity measures based on the new node representations and iteratively updates the weights of edges until convergence.In addition,the relationship between the proposed method and the related depth graph model is analyzed theoretically in this paper,and the experimental results demonstrate the superiority of the proposed model.(2)In this paper,the graph topology is studied in depth from the perspective of edge distribution,and the role of edge distribution for neighborhood aggregation is demonstrated through theoretical analysis and experiments.Based on the theoretical and experimental approaches,a new method(GCN-IED)is proposed to improve the neighborhood aggregation process in terms of both direct and hidden edges,which mainly includes updating the graph topology with local neighborhood information and fusing multi-order neighborhood information.A large number of experiments are also conducted to demonstrate the effectiveness of the proposed method,especially on heterophilous graphs.In conclusion,the research in this paper investigates the problem of over-smoothing in graph convolutional networks and the problem of neighborhood aggregation in heterophilous graphs,enhances the graph structure and improves the neighborhood aggregation process from several aspects,and demonstrates the effectiveness of the method from a large number of experiments.The research in this paper provides new methods and techniques for graph data node classification,which has high application value in social networks,community discovery and other fields.