| Graph data often used to describe the relationship between things.Many things in real life,such as online social networks,protein interaction networks,etc.,can be abstractly described as graph data composed of individuals and their connection relationships.With the tremendous success of deep learning in classification problems,it has become a research focus to transfer the successful experience of deep learning to graph data classification task.Classification tasks based on graph data can generally be divided into node-level classification tasks and graph-level classification tasks.Among them,classification tasks are the most common as an important data mining method.For example:treat users in online social networks as a node,classify users by extracting graph features of the nodes to the subreddit to which they belong,such tasks are node-level classification;treat the protein network as a graph,classify the protein as enzyme or non-enzyme by extracting the graph features of the networks,such tasks are graph-level classification tasks.However,when dealing with these problems,the existing graph neural networks usually have problems such as incomplete feature extraction,difficulty in training deep networks,over-smooth and low classification accuracy.In order to solve these problems,this article has done the following work:(1)In view of the problems that the existing deep graph neural network does not consider the layer features in the graph classification task,difficulty in training deep network,and the low classification accuracy,An hierarchical graph classification method based on importance pooling is proposed.Taking the “intra layer-inter layer joint feature extraction structure” as the core and mainly includes two parts of works: intra-layer feature extraction module and inter-layer feature extraction module.First,using the graph convolution neural network to extract the initial features of the graph data,the graph with new features is obtained.Second,the importance pooling method is used to sample the important nodes in the graph and form a thumbnail containing most of the features and properties of the original graph.Then,by stacking the graph convolution layer and the graph pooling layer to extract deep features of the graph data.Finally,forming the feature matrix of the thumbnails output by each graph pooling layer as sequence data,and input to the recurrent neural network,and use the logic gate structure to assign appropriate weights and receptive fields to the output of the feature extraction module in each layer,so that the final results of the model are adaptively integrated into the feature information of different layers.The experimental results show that under a reasonable time complexity,our methodcan accelerate the convergence of the loss function,the classification accuracy on the benchmark graph classification datasets D&D,PROTEINS,NCI1 and MUTAG has been improved to varying degrees.(2)Aiming at the problem of over-smoothing in the node classification task of existing deep graph neural networks,a multi-scale node classification method based on subgraph partition is proposed.First,a method based on subgraph partitioning is used to preprocess the graph data.By changing the network structure in the graph,the low-yield feature aggregation work is effectively avoided,to a certain extent,the problem of over-smooth caused by redundant neighborhood search is suppressed.Then,a Graph-Inception network structure is designed,through the combination of convolution kernels of different sizes to extract the neighborhood feature information of the target node at multiple scales,so as to achieve the equivalent of the depth expansion of the graph neural network,to a certain extent,over-smooth caused by too many model layers.Experimental results show that this method can effectively suppress the over-smoothing phenomenon in graph neural networks,the classification accuracy on the benchmark node classification datasets PPI,REDDIT and Amazon has been improved to varying degrees. |
PDF Full Text Request