关键词: 图表示学习, 图神经网络, 数据增强, 自监督学习, 图对比学习, 节点分类

Abstract: Graph Neural Network (GNN) is an effective graph representation for processing graph structure data. However, the performance of GNN in practical applications was limited by the problem of missing information. On the one hand, the graph structure was usually sparse, making it difficult for the model to adequately learn node features. On the other hand, model training was limited by the fact that labels data, on which supervised learning relies, were often scarce, making it difficult to obtain robust node representations. To address these problems, a Subgraph-aware Contrastive Learning with Data Augmentation (SCLDA) model is proposed. Firstly, the relationship scores between nodes are derived by learning the original graph through link prediction, and the edges with the highest scores are added to the original graph to generate the enhanced graph. Secondly, local subgraphs are sampled using target nodes for the original and enhanced graphs, respectively, and the target nodes of subgraphs are input to the shared GNN encoder to generate target node embeddings at the subgraph level. Finally, mutual information between similar instances are maximized based on contrastive learning of the target nodes from the two perspective subgraphs. Experiments on node classification on six public datasets Cora, Citeseer, Pubmed, Cora_ML, DBLP, and Photo show that SCLDA improves the accuracy over the traditional GCN model by about 4.4%, 6.3%, 4.5%, 7.0%, 13.2% and 9.3%, respectively.

Key words: graph representation learning, graph neural network, data augmentation, self-supervised learning, graph contrastive learning, node classification