Abstract: The universally accepted technique to address the issues of computational complexity and high spatial complexity of adjacency matrix due to non-Euclidean spatiality of graph data is to use graph embedding methods to map high-dimensional heterogeneous information,such as graph topology and node attributes,to dense vector space.In this paper,based on the analysis of the problems of the classical graph auto-encoder model GAE(graph auto-encoder) and VGAE(variational graph auto-encoder),we try to improve the graph embedding method based on graph auto-encoder from three aspects:encoder,decoder and loss function,and propose a graph auto-encoder model NLGAE based on the improved network structure and loss function.First,in the model structure design,on the one hand,the stacked graph convolutional layers in the encoder are inverted to solve the problem of lack of flexibility and insufficient expressiveness of the non-reference decoder in GAE and VGAE,on the other hand,the graph convolutional network GAT is introduced to solve the problem of solidifying the weight coefficients between nodes by introducing the attention mechanism.Second,both the graph structure and the node feature information could be taken into account by the redesigned loss function.The comparative experimental results show that,as an unsupervised model,the proposed NLGAE can learn high-quality node embedding features and outperform not only traditional unsupervised models DeepWalk,GAE,GrpahMAE,GATE,etc.in node classification tasks,but also supervised graph neural network models such as GAT and GCN in the case of selecting an appropriate classification model.

Key words: Graph representation learning, Graph auto-encoder, Attention mechanism, Node classification