Research On The Over-smoothing Problem In Graph Node Classificatio

As the understanding of Graph Neural Network(GNN)has become deeper,many limitations of existing GNNs have been identified.Some of these limitations are inherited from the traditional Deep Neural Network(DNN),while others are specific to GNNs,such as increasing the number of layers in a GNN leads to a significant performance degradation.Some recent work attributes this performance degradation to the over-smoothing problem.Each layer of a graph convolution network is such that each node aggregates as much information as possible from its surrounding nodes,and averages the surrounding node features to obtain new features for each node,and performing multiple graph convolution operations causes the values of the connected nodes in the graph to converge to the same value.Therefore,when multiple propagation is performed,node representations from different classes will become indistinguishable,which makes GCN’s capability very limited in some tasks,especially for graph structures containing less information.Thus,it is necessary to increase the number of convolutional layers to obtain richer node information by exploiting the higher-order neighbor information of nodes.Designing a GCN model to effectively prevent over-smoothing and achieve better results in deep network structures is still a pressing research problem.The research in this paper aims to solve the over-smoothing problem in GNNs,including the following three approaches.The research in this paper aims to solve the over-smoothing problem in GNNs,including the following three approaches.First,an Adaptive Graph Convolutional Networks based on Decouple and Residuals to relieve over-smoothing(ADR-GCN)is proposed.Specifically,ADR-GCN first uses a self-encoder to reduce the dimensionality of the input node features to generate subsequent node embeddings for propagation.Then the aggregation of nodes is guided by the adjacency matrix,i.e.,the deeper the node embedding aggregates the larger the range of neighborhood information.To preserve the original feature information of the nodes,an initial residual connection is used to ensure the retention of local information.Since the node embeddings of different layers contain multi-scale information,an adaptive aggregation mechanism is designed to aggregate the node embeddings of different layers in order to fully integrate the multi-scale information.Various experimental results show that ADR-GCN has high accuracy for node classification and can better mitigate over-smoothing.Second,a Graph Neural Networks via node-dependent Local Smoothing and Feature Complementation(LSFC-GNN)is proposed.First,LSFC-GNN controls the smoothing of each node by setting a node-specific smoothing iteration.Specifically,LSFC-GNNN calculates the influence score based on the adjacency matrix and selects the number of iterations by setting a threshold on the score.Once selected,the number of iterations can be applied to feature smoothing.In addition,an auxiliary task of feature complementation is designed,which aims to help GNN learn better attribute information.The results of various experiments show that LSFC-GNN has high accuracy on the node classification task and can better mitigate over-smoothing.Finally,a global structure-aware and feature similarity-preserving graph convolutional neural networks(GSA-FSP)is proposed.Specifically,the corresponding KNN structure matrix is computed based on the given node features,and at this time the KNN structure matrix has feature similarity information.Then the global structure matrix is obtained by Kmeans calculation,which takes into account the global structure information of the nodes.The node adjacency matrix initially given is combined with the above two matrices to obtain the enhanced adjacency matrix for propagation,which is used to guide the subsequent node aggregation update.In addition,node contrast loss is designed to keep unrelated nodes away from each other in the representation space.Experiments are conducted on three benchmark datasets,and the results show that GSA-FSP achieves advanced results and can better mitigate over-smoothing compared to existing baseline methods.