Research And Application Of Node Classification And Graph Classification Based On Graph Convolution Network

Graph convolution network has attracted much attention because of its ability to aggregate and transform information within node’s neighbors to generate a node representation.As an important criterion for evaluating model embedding effect,classification task is an important research direction in graph analysis task.Due to the complex connections and more noise in the social network,the existing models have the problems of dependency graph structure and local structure being ignored,resulting in the inability to obtain effective feature representation.In response to these problems,this paper conducts research on node classification and graph classification methods based on graph convolutional networks.The detailed work is as follows:(1)To address the problems that existing graph convolutional networks rely on graph structure information and a single way of node feature propagation in node classification tasks,a semi-supervised node classification method that preserves feature similarity is proposed.According to the cosine similarity between nodes,the first k similar nodes are connected to form the feature graph,which is used to preserve the feature similarity of nodes.The introduction of a mixed graph convolution module to adaptively integrate the adjacency matrix of the initial and feature graphs,which can compensate to some extent for the information lost in the aggregation process.The multi-channel mechanism allows node features to be propagated over the three graph structures and adaptively fuses the embedding of the three modules using an attention mechanism,which can effectively reduce the dependence of the model on the graph structure information.The experimental results show high effectiveness on five different paper citation networks and social network datasets compared to other baselines.(2)To address the problems of poor representation of the entire graph in existing graph convolutional networks in graph classification and the Transformer’s inability to obtain the structural information of the graph,a graph classification method based on Transformer and GCN is proposed.A global encoding module is introduced to add degree encoding to the feature matrix of the input nodes,so that nodes with more neighbors are given more weight in the attention mechanism.The local aggregation module aggregates the features of the encoded neighbor nodes,preserving the local structure of the graph and obtaining a better representation of the nodes.In addition,a global pooling structure is used to obtain a feature representation of the whole graph using the READOUT function.The method takes advantage of the complementary features of the two models for modelling,with good results on different social network and bioinformatics datasets.(3)Based on the node classification model and graph classification model proposed in this paper,the relational network classification system is designed,the main functions and tests of the system are realized.The key modules of the system include the node classification function module and the graph classification function module.A semi-supervised node classification method that preserves feature similarity implements the node classification function,and a graph classification method based on Transformer and GCN implements graph classification functions.