Research On Node Classification Algorithm Based On Graph Neural Network

As a kind of data structure that can picture the relationship between things,the graph can abstract the data in the real scene into a graph,visually represent the complex relationship between the data,and analyze the data more conveniently.Node classification utilizes the labeled nodes on the graph to predict the labels of unlabeled nodes and is the key in the field of graph data analysis.With the development of deep learning,the use of graph neural networks for graph data classification has become an important research hotspot.However,there are still some problems to be solved,such as insufficient utilization of node features,over-smoothing of deep graph neural networks,and low accuracy of multi-label classification.Because of the above problems,this paper studies the node classification algorithm based on graph neural networks,the specific research content is as follows:(1)Aiming at the problems of the small receptive field and insufficient node representation in existing node classification tasks,a structure-feature aware graph neural network model is proposed.Using the hierarchical structure,by combining the pooling graph and the unpooling graph,the local information of the graph and the global structural information of the graph are fully captured,so that the information available for the node classification task is more sufficient and the accuracy rate is higher.Experiments show that the proposed graph node classification method has an outstanding performance and improved performance compared with the baseline method.(2)Aiming at the problem of over-smoothing when using a deep graph neural network for node classification,an improved graph attention network method based on clustering is proposed.First,the nodes are divided into different clusters to form multiple subgraphs,and the neighbor nodes of the target node are selected in each subgraph,and then the multihop method is used in the attention network of each layer to aggregate the multi-order neighbor information of the nodes to alleviate the oversmoothing problem.The experimental results show that the proposed node classification method is compared with the baseline methods and obtains a better node classification effect.(3)Aiming at the problem that the node attribute information,label information and graph topology information cannot be used at the same time when classifying multi-label nodes,a label-aware multi-label node classification method is proposed.First,the label attention module is used to capture the node-label dependencies.Second,according to the idea of reinforcement learning,an agent is assigned to each label,which is used to travel on the multi-label graph to capture the relationship between nodes and the structure of the graph.The experimental results show that the proposed method achieves an outstanding node classification result than the baseline methods.