Research On Graph Classification Technology Based On Graph Neural Network

As a general-purpose data structure,graph data has been widely used in various real-world areas.Graph data is non-Euclidean data,it is not clearly arranged like Euclidean data,and the number of neighbouring nodes in graph data is not fixed and the order between nodes may be different,making it difficult to define Euclidean distances.It is therefore necessary to embed graph data into a suitable Euclidean structure,and thus researchers have extended neural networks to graph data and proposed graph neural networks.The aim is to use the powerful feature modelling capability of graph neural networks to solve graph mining problems.In recent years,several deep learning methods have been explored for processing graph data,and significant breakthroughs have been achieved in node classification and link prediction.However,there is still much room for improvement in graph classification tasks.For example,some graph classification methods assume that all nodes have the same importance,which can cause the different importance levels among nodes to be ignored.In fact,some nodes in a graph often have more critical and important information,such as the core atoms of a biomolecular structure or the central nodes of a traffic hub in a transportation network,and the way these nodes are represented is to some extent more important.Moreover,most of the current graph classification methods do not perform well in large-scale complex graphs,and cannot quickly form embeddings of nodes in the graph,which can easily lose important features of the graph.In this paper,we combine inductive graph convolutional networks with selfattentive pooling to propose a new graph classification network framework called graph classification network framework based on inductive learning graph convolution and self-attentive pooling to address the above problems.Firstly,to address the problem that current methods are not efficient in forming embeddings of nodes,the convolutional approach of the inductive learning Graph SAGE model is used in the graph convolutional layer of the overall network framework,so that the overall framework can form embeddings of nodes in the graph more quickly.Secondly,to address the problem that the current method tends to ignore the difference in importance of nodes in the graph,a graph isomorphic aggregation strategy is used in the aggregation method of the Graph SAGE model,so that it can retain the graph features more accurately.Then,to address the problem that the current method tends to lose important features of the graph,an improved self-attentive scoring method is used in the pooling layer to avoid the loss of important features.Finally,a hierarchical framework suitable for extracting large-scale graph information is adopted in the overall architecture,combining the graph convolution layer,the graph pooling layer and the graph readout layer for the downstream graph classification task.Experimental results show that the network framework provides a very significant improvement over other graph classification methods with the same common dataset.