Research On Node Classification Based On Graph/Hypergraph Neural Networks

Graph node classification is widely used in social networks,e-commerce and disease prediction.The complexity of graph structure brings challenges to the application of existing classification algorithms,so it is great practical significance to explore efficient classification algorithms.This paper mainly studies the semi-supervised classification method of nodes based on graph convolutional neural network and hypergraph neural network.In the node classification method of the graph convolutional neural network,the initial graph structure often has noise,which is directly sent into the network model for training,and the classification accuracy of the model will be affected.Therefore,this article first improves on the existing graph neural network architecture,and designs a measurement method that combines graph structure and node attribute information to reduce the noise in the underlying graph structure.In addition,when considering the mining of information on the graph structure that characterizes the paired relationship,there are problems that the global feature information is incomplete and the high-order relationship between objects is ignored.By extending the graph to the hypergraph and improving on the dynamic hypergraph neural network,design a dynamic hypergraph neural network(RWS-Dynamic Hypergraph Neural Networks,RWS-DHGNN)that combines graph random walks and skip connections.To obtain global feature information and high-order correlation information between data.The work and research results of this paper are as follows:1.Node classification network based on the combination of graph structure and node attribute information.Aiming at the Iterative Deep Graph Learning(IDGL)network only uses node attribute information when measuring the similarity between nodes,ignoring the geometric topological structure relationship,and design a measurement method that combines graph structure and node attribute information.By jointly considering the underlying geometric structure of the graph and the node attribute information,an adjacency relationship is added to measure the structural similarity between nodes,so that the similarity calculation between nodes is more accurate,and the learned graph structure noise is reduced.Experiments prove that the proposed method has a good classification effect compared with methods such as IDGL.2.The dynamic hypergraph neural network of joint graph random walk and skip connection.Compared with graphs that can only model second-order relationships,hypergraphs can more accurately represent higher-order relationships between objects.In view of the difficulty of hypergraph neural network to extract node features with high correlation outside the node's direct neighborhood,which leads to the problem of incomplete global feature information,a dynamic hypergraph neural network joint graph random walk and skip connection(RWS-DHGNN)is designed.The network combines the ideas of dynamic hypergraph neural network,graph random walk and skip connection.First of all,on the basis of DHGNN,a graph random walk is introduced to effectively obtain the features of nodes with high correlation outside the immediate neighborhood.At the same time,skip connections are added at the vertex convolution of the hypergraph to form a residual structure,which improves the classification performance of the model.The RWS-DHGNN model effectively exerts the advantages of graph structure and hypergraph structure.Experimental results show that compared with GCN,HGNN and DHGNN,the proposed network effectively improves the accuracy of node classification.In summary,this article mainly studies the related technologies and algorithms of graph convolutional neural networks and hypergraph neural networks,and improves the accuracy of node classification from two aspects: reducing graph structure noise and increasing global feature information.