Research On Semi-supervised Node Classification Based On Graph Convolution Neural Network

The classification of graph nodes is widely used in social networks,biomedical Science,e-commerce and other fields,and has always been a hot topic of research by scholars today.The increase in the scale of graph nodes has brought huge challenges to the existing classification algorithms.It is very necessary to explore efficient classification algorithms to meet people's urgent practical application needs.In this paper,we study the key technologies of semi-supervised node classification of graph convolution neural network.First,we add graph pooling layer and random walk to graph convolution neural network,which effectively improves the accuracy of the model.Secondly,to improve the efficiency of graph convolution neural networks,a new adaptive sampling fast graph convolution neural network is designed,which is suitable for larger graph node classification.The work and research results of this paper are as follows:1.Aiming at the problem that graph convolution neural network(Graph Convolution neural network,GCN)has low classification accuracy and shallow model,and cannot fully extract graph node features,a deep pooled dual graph neural network is designed.First,the graph convolution operator is constructed using the principle of graph convolution network;then the convolution layer is used to retain the local geometric multi-scale structure of the designed graph pooling operation and refinement operation to form a deep pooling neural network to extract deeper semantics feature information;Finally,collaborative training is carried out through random walks and deep pooling networks,and the labeled data is effectively propagated to the entire graph to solve the localization problem of the convolution kernel.Experimental results show that the network model improves the classification accuracy compared with the existing methods,and the effect is more obvious when a small amount of data is marked.2.Aiming at the problems that graph convolution neural network takes a long time to train when classifying large-scale graph nodes,and the server consumes a lot of memory,this paper designs a fast graph convolution neural network model with adaptive importance sampling.First,the importance nodes are sampled from top to bottom,and the bottom nodes are sampled as a whole to avoid excessive expansion of the neighborhood and achieve the effect of reducing running memory;secondly,variance reduction is used in the forward propagation process of the network Accelerate training;finally,the extracted hidden layer features and the importance-sampled fast graph convolution stacking features are enriched by random walk graph convolution.Experimental results show that the network has good classification performance on large-scale graph data,and at the same time can also reduce the server's operating memory overhead to a certain extent.