Research On Semi-supervised Node Classification Based On Deep Graph Convolutional Networks

Graph convolutional networks are widely used in semi-supervised node classification tasks based on graph-structured data.However,as the scale of graph data continues to expand,the connections between nodes become more and more complex.and only simple or shallow graph convolutional network models can no longer meet the requirements of the task.Therefore,more complex and deep graph convolutional network models are needed to handle large-scale graph data for more accurate and efficient node classification.However,with the deepening of the number of model layers,a series of problems have emerged,such as over-smoothing,loss of initial information,long training time and reduced accuracy,etc.These problems hinder the further development of deep graph convolutional networks.In order to deeply study and solve the over-smoothing problem in deep graph convolutional networks,the paper first proposes a method to measure the degree of over-smoothness of nodes.In order to ensure the consistency between the nodes in the update process and the initial features of the nodes,an initial residual depth map convolution network based on hierarchical feature constraints is designed,and the oversmoothing problem is alleviated by optimizing the over-smoothing index.Further focusing on the over-smoothness of different nodes in the graph convolutional network layer in the model,an adaptive initial residual depth graph convolutional network based on node smoothness is proposed.This method can assign different initial features according to the degree of over-smoothing of the nodes in the graph,so as to further alleviate the over-smoothing problem.Specific work includes:(1)Aiming at the over-smoothing problem in deep graph convolutional networks,a new measurement method using the difference before and after node update to evaluate the degree of node over-smoothness is proposed,which reduces the time for calculating node smoothness compared with existing methods.Some methods to alleviate the oversmoothing problem are further analyzed,including optimizing the oversmoothing index and adding initial residual connections,etc.(2)Aiming at the problem that initial information is lost when nodes propagate in deep graph convolutional networks,an initial residual depth graph convolutional network based on hierarchical feature constraints is proposed.This method introduces similarity constraints and difference constraints to limit the degree of deviation between the current layer features and the initial features,and the similarity between the current layer features and the previous layer features.By optimizing these two constraints in the loss function,it can be ensured that the node features will not deviate excessively from the initial features after multiple iterative updates,and remain sufficiently different from the last update to obtain more useful classification features.(3)In view of the fact that the nodes in the graph have different degrees of oversmoothing during the update process,we propose an adaptive initial residual depth graph convolutional network based on node smoothness.Experiments show that nodes with different degrees of over-smoothing should be processed separately,that is,nodes with more severe over-smoothing need more initial information to alleviate its oversmoothing problem.Therefore,this method assigns an appropriate proportion of initial features to nodes according to their smoothness to alleviate the over-smoothing problem and enable the construction of deeper graph convolutional networks.