Graph Neural Network Method Based On Graph Structure Learning And Its Application In Node Classification

With the development of the information age,many types of graph data are generated in real life,and mining the potential connections between data through graph neural networks has become a hot research topic.However,most traditional graph neural networks smooth the node features in the process of node feature aggregation,which tends to destroy the similarity of nodes.And as the number of network layers increases,the over-smoothing problem in the network becomes more prominent,resulting in a significant decline in network performance.On the basis of the existing work,this dissertation focuses on the graph neural network method based on graph structure learning and proposes a series of effective network structures to achieve advanced node classification performance,and the main works are as follows:1.A deep graph neural network based on graph structure transformation and dynamic feature update is proposed.The network is composed of a deep local branch and a global branch for feature extraction.First,different adjacency matrices are generated in the deep local branch by Drop Edge to Increase the diversity of graph structure.The adjacency matrix and node features are input into the adaptive filter to learn the deep features.Subsequently,an adaptive skip connection memory block is proposed to dynamically store the deep features of different layers.Finally,the deep features and global features are fused to complete the node classification task.Experiments show that the network depth of the proposed method can reach 128 layers and achieve the expected classification accuracy.2.A graph neural network based on random reconstructed graph structure is designed,which solves the problem of node similarity destruction by reconstructing the graph structure.First,input features are enhanced by random feature transformation to randomly reserved nodes to generate random features.Then,the original graph is fused with the k-nearest neighbor graph according to a score vector calculated by the random feature,in order to obtain the random graph structure adaptively.Finally,the model extracts multi-branch information from the initial nodes,and supplements more initial information for the reconstructed graph with the deepening of the number of layers.Experiments show that the proposed method achieves better performance on the node classification task.3.A graph neural network based on graph structure transformation is constructed to enhance the initial information,which effectively enhances the initial information through local similarity branch and self-attention branch.Specifically,the local similarity branch is to perform similarity metric learning and random feature transformation on the initial nodes,and the random features carry out feature propagation according to the knearest neighbor graph structure,in order to maintain the node similarity to the greatest extent.The self-attention branch computes the internal correlation of features according to the graph structure,capturing key node features.Finally,two learnable branches adaptively supplement similar node information and attentional node information to the backbone model.The experimental results show that this method can enhance the role of initial information in the network,thereby improving the performance of the model node classification task.