Graph Neural Network With Mixed High And Low Order Information And Its Application In Semi-supervised Classification

As one of the most effective neural networks for processing graph data,graph neural networks have been involved in all aspects of life,from search recommendations to e-commerce shopping.The graph convolutional neural network transfers the convolution operation to the graph data with the help of the Laplacian matrix of the graph,and it can complete related tasks such as the classification of nodes in the graph,the prediction of edges,and the classification of graphs.Therefore,the graph convolutional neural network can also attracting more and more scholars to conduct in-depth research.However,the current graph convolutional neural network aggregates information only for the first-order neighbors directly connected to the target node,and does not fully mine the high-order neighbors that are also rich in information.Therefore,in view of the problems of the current graph convolutional neural network,this thesis proposes a graph convolutional neural network that can consider both high-order and low-order neighbor information.The main work content can be divided into the following two points:(1)Graph neural networks are widely used in various deep learning tasks because of their powerful ability to process graph data.Among them,the popular graph convolutional network simply aggregates the features of neighbor nodes,and this oversimplified operation limits the model learning ability.Therefore,this thesis first proposes a graph convolutional network algorithm that mixes high-order and low-order neighbor information.For the new graph convolutional network layer,through power multiplication,the information of multiple adjacency matrices can be synthesized,so that the information of higher-order neighbors can be learned information.In view of the difference in information rich in each order of neighbors,a learnable parameter is set for each order of neighbors.Compared with traditional graph convolutional networks,the computational complexity of the algorithm and the actual memory requirements have not increased further.For graph convolution Network,the thesis also proposed a new convolution operator,so that the graph convolution algorithm can be better suitable for the current semi-supervised classification tasks.In several standard semi-supervised citation datasets,compared with other current graph convolution algorithms,experiments show that the effect of the algorithm proposed in this thesis is better.(2)By introducing the attention mechanism,the graph attention network can assign different weights to different neighbor nodes,and does not require complex matrix operations.However,the current attention mechanism can only consider the nodes directly connected to the central node,ignoring the multi-hop neighbors that can provide important information.Therefore,this thesis designs a graph attention network based on a multi-hop connection mechanism,which can mix the current graph attention network.By adding multi-hop information to the attention mechanism,the receptive field of each layer is increased,making the network more robust,and completing the semi-supervised classification task based on graph convolution.In several standard graph datasets,the experimental results show that the hybrid graph attention algorithm proposed in this thesis is better than the current graph attention network algorithm.