Research On Graph Classification Algorithm Based On Graph Convolutional Network

Many data in the real world,such as social networks,compounds,biological proteins,etc.,can be usually represented by graph structures,which can describe complex interrelationships between objects or organizations.Therefore,the analysis of such data can be abstracted as a graph classification problem.However,the graph data is a kind of non-Euclidean structure data,which is not as neatly arranged as Euclidean structure data.For different nodes in the graph data,the number of neighbor nodes is not fixed.At the same time,because the order of the nodes in each graph structure may be different,it is difficult to define the Euclidean distance.Thus,the graph data needs to be embedded into a suitable Euclidean vector space to solve the problem of graph classification.Recently,many deep learning-based methods have emerged to process graph-structured data,with great success in the tasks of node classification and link prediction.However,there are still many shortcomings when dealing with graph classification tasks.This paper focuses on two types of methods,spatial convolution and attention-based mechanisms,to analyze and study graph classification methods based on graph convolutional networks.The specific research contents are as follows:(1)Spatial convolution method based on structural features of significant nodes for graph classification.Existing spatial convolution methods rely on a single indicator or the structural characteristics of the nodes to measure the significance of nodes when selecting the central node sequence,which will lead to the loss of some important nodes.Therefore,a comprehensive weighting method based on multiple indicators to evaluate the importance of the nodes is proposed.Afterwards,in connection with the singleness of the node features input to the convolutional layer in the spatial convolution method,we propose to select multiple node features from both local and global aspects,retaining more graph structure information.Finally,the channel connection layer is introduced and added after the convolution layer,so that the classification result of the entire graph can be determined according to the classification of the neighborhood graph.Experimental results on 7 benchmark datasets show that the proposed method is superior to some classic graph kernels and deep learning-based spatial convolution models.(2)Graph classification based on node-level and subgraph-level attention mechanisms.Aiming at the problem that the existing attention-based methods will ignore some node information or existing edge information,this paper enhances the initial node feature matrix and aggregates the neighbor information of nodes.Then at the node level,the attention mechanism is applied to automatically learn the importance of each vertex to select the sequence of nodes to generate the subgraph structure.After that,a graph convolution layer is used to learn the subgraph representation and the attention mechanism is applied at the subgraph level to learn the significance of different generated subgraphs.Finally,all the subgraph representations learned by the graph convolution layer are weighted and summed with the corresponding importance coefficients to obtain a graph representation.At the same time,multiple weight initialization queries can generate a multi-scale graph representation.Experimental results on 4 datasets with node labels prove that the final graph representation retains the rich information of the graph and improves the classification accuracy.