Neural Network Algorithm For Node Classification

Graph-structured data widely exists in the real world.Citation network,social network and transportation network are typical graph-structured data.Node classification is one of the most common and important tasks in the field of graphs.Graph convolutional network is an extension of convolutional neural network in the field of graph,which solves the problem that traditional convolutional neural network cannot handle non-Euclidean structural data,and has been widely used in processing graph and network data.Node features and topology structure are both important information in graph-structured data.However,some recent research experiments show that existing graph convolutional networks cannot fusion node features and topology structure well.Specifically,due to the sparsity and noise,similar nodes may be far apart whereas different types of nodes are directly connected,but graph convolutional networks does not consider these phenomena.In addition,most current graph neural networks only learn the features of nodes and often ignore the attention to edges.Researchers have found that learning edges can better obtain graph information.Finally,traditional graph neural networks face several unavoidable problems,such as over-smoothing and so on.In view of the above findings,this thesis designs two neural network algorithms that can improve the fusion ability of node features and topology structures.The specific methods are as follows:(1)For existing graph convolutional networks that cannot well fuse node features and topology,propose a cross-stitch unit based graph convolutional network model for node classification.In order to more effectively utilize the node features and topology structure in graph structure data,the cross-stitch unit used in traditional convolutional neural networks is introduced into the graph convolutional network model.The cross-stitch network spread specific information or knowledge between node features and topology structure,and obtains consistent learned representation by integrating information of node features and topology structure at the same time.Therefore,the cross-stitch network can capture various channel information in all graph through multiple channels.In addition,attention mechanism is introduced into the model,which aims to adaptively learn the weights of channel embeddings.Furthermore,feature selection methods are utilized to denoise the data and alleviate sparsity.(2)Aiming at the unavoidable over-smoothing problem faced by traditional graph neural networks,this thesis proposes a sequential neural network model based on collaborative attention mechanism for node classification.First,two Graph Transformer architectures are designed to learn the topological graph and feature graph respectively,and the node features and topological structures in the graph are put into two different Graph Transformer Layers,which are operated by the self-attention mechanism respectively.Then,the obtained node embedding results are computed through a co-attention mechanism and fed into a feed-forward neural network.The above process is regarded as a co-attention operation,and the fusion information of node features and topology structure in the graph structure data can be fully extracted after repeated training for many times.Finally,two hyperparameters are used to adaptively learn the weight of each loss through the attention mechanism,and node classification is learned by calculating the distribution probability of node features and calculating the cross-entropy loss between the true label and the predicted label.It is worth noting that in order to better learn the graph information,the algorithm also learns the edges.In order to verify the validity of the model in this thesis,experiments on six benchmark data sets and comprehensive analysis of the experiments have been carried out.On different evaluation indexes,the proposed method has achieved good results.