Research On Graph Classification Algorithm Based On Graph Convolutional Neural Network

In modern information society,data structure is complex and data volume is huge,while graph data can describe complex network models via object-to-object connections,which is why graph data are used in various social networks,like social networks,protein molecular structures,and citation networks.As an abstract representation of real-time networks in the domains of graph learning,how to capture node feature information and topological structure information in graph data effectively is a key point.With the remarkable achievements of deep learning in various areas,it is now a hot research topic to apply it into graph learning domain.The graph-level classification task is a major component of graph learning field,yet the existing algorithm of graph-level classification task have some problems such as insufficient feature extraction,inability on capturing the bidirectional dependencies between read graphs,lack of node information,and omitting the structural differences in graph data.Aiming to address the foregoing problems,this paper develops the following research:(1)Aiming at the problems of insufficient feature extraction and failure to capture the bidirectional dependencies between the readout graphs in the graph classification algorithm,this thesis proposes a hybrid hierarchical graph classification algorithm based on BiLSTM.The algorithm evaluates the importance of nodes from the point of view of nodes and structure,and then fuses feature information to effectively increase the diversity of node evaluation.In addition,because there is a sequential dependency between the readout graph and the readout graph,this thesis uses BiLSTM to capture the forward-backward dependency between the readout graphs,so as to extract rich feature information at a deeper level,and then obtain a more representative graph feature representation.(2)Aiming at the problems of missing node information and ignoring structural differences in graph data in the graph classification algorithm,this thesis proposes a graph classification algorithm based on graph isomorphic aggregation and multi-level fusion modules.The algorithm uses the graph isomorphism aggregation function to aggregate the feature information of the surrounding neighbor nodes.It can identify the structural differences in the graph data.In addition,in the previous pooling operation,when the node importance is scored,the feature information is fused by weighted sum,and this thesis adopts the maximum value method for fusion,taking out the maximum value of the node in two different importances as the importance score of the node,so that more important nodes can be saved,and finally a multi-level fusion module is proposed to avoid due to the problem of node information loss in the pooling operation,and the last node contains information of all layers,solves the problem of unbalanced distribution of node features due to different graph topological structures.Finally,through various experimental analysis,the feasibility and effectiveness of the two graph classification algorithm are confirmed,and the classification accuracy on the D&D,PROTEINS and MUTAG data sets is improved compared with the comparison algorithm.