Research On Smoothness Of Graph Convolutional Networks

In recent years,with the further research of data mining on complex networks,Graph Convolutional Network(GCN)has become an effective tool for analyzing graph structure data,which has achieved good results in node classification,link prediction,graph classification and other tasks.Graph convolution operation is a special form of Laplacian smoothing.The smoothness research of GCN is of great significance to further study graph convolution model.This thesis takes the smoothness of graph convolutional network as the theme and explores how to enhance the expression ability of graph convolutional network from local and global perspectives respectively.From the perspective of local smoothness,GCN completes feature update of nodes through feature aggregation of first-order neighbor information of nodes,but only the first-order subgraph will lose part of higher-order structure information of the network.In order to capture richer information of graph topology,MS-GCNs is proposed.MS-GCNs identifies the motif structure in the network and combines the node motifstructure information to learn the weight of message aggregation,so that the nodes can aggregate higher-order network information to enhance the local smoothness of graph convolution and improve the expression ability of GCN model.This thesis improves the most representative three graph convolutional network models,taking node classification as the basic task,and achieves good experimental results in real networks.From the perspective of global smoothness,although Graph Convolutional Network has achieved good results in many learning tasks,its essence is still an information aggregation process of first-order neighbors.As the number of network layers increases,node features in the network tend to be consistent,resulting in global over-smoothness.In order to solve these problems,the Adaptive Connection for Training Deeper Graph Convolutional Networks(ACGCN)model is proposed in this paper.ACGCN uses the initial residual based on the existing residual network model and allows each layer of network to adaptively connect to all the previous layers.This thesis carries out node classification experiments on different types of real networks,and the experimental results show that ACGCN can not only effectively solve the over-smoothing problem,but also improve the expression ability of the model with the stacking of network layers.In addition,compared with the existing models based on residual connection,the superiority of the model is proved theoretically and experimentally.