Research On Node Classification Based On Graph Convolutional Neural Network

Graph neural network,a significant approach to deal with large scale graph-structured data,can potentially capture the topological structure information and feature information.And,node classification is very common in the real world: training a model to learn in which class a node belongs.In this paper,a new graph embedding algorithm for node classification in unsupervised cases and a new node embedding algorithm in heterogeneous graphs are proposed.The main work of this paper is as follows:(1)the basic concepts and related definitions of graph are introduced,and then different types of graph structure data are introduced with examples.In addition,network embedding algorithms based on matrix decomposition and network embedding algorithms based on random walks are studied,and their characteristics,advantages and disadvantages are discussed.(2)three classical deep graph neural networks are discussed,including Graph Convolution Network(GCN),Graph Attention Network(GAT)and Fast GCN.These three models are compared and tested on three classical node classification benchmark datasets(i.e.,Cora,Citeseer and Pubmed).They lay the foundation for subsequent research and improvement and will be used as the benchmark model in the field of graph neural network.(3)a new unsupervised hierarchical graph embedding model based on mutual information maximization is proposed,which named Info HGraph,its key idea to be success is to maximize the mutual information between global graph embedding representation and local node embedding representation as well as the mutual information among hierarchical representations of thr graph.Info HGraph can be divided into three modules: encoder module,graph pooling module and discriminator module.In the case of unsupervised learning,the node classification accuracy of Info HGraph increased by 23.3%,69.9% and 20.1% on CORA,Citeseer and Pubmed respectively,compared to Deepwalk.Compared with Deep Graph Infomax(DGI),the classification accuracy on the three datasets are increased by 0.72%,2.2%,and 2.1%,respectively.(4)a new model is proposed for learning heterogeneous graph embedding representation which named AHGraph.AHGraph adopts a hierarchical attention structure,including type-level attention,node-level attention and semantic-level attention.The hierarchical attention mechanism can learn the heterogeneity of different types information so that to obtain node embeddings involving specific semantics for a specific task.In the node classification experiment on ACM dataset,compared with the Heterogeneous Graph Attention Network(HAN),its micro-F1 score and macro-F1 score are increased by 1.94% and 1.86%,respectively.And its Micro F1 score and Macro F1 score on IMDB dataset increased by 5.56% and 31.74%,respectively.