Global Augmentation and Local-aware Multi-view Graph Contrastive Learning

Graph Neural Networks (GNNs) have been shown to be effective graph representations for processing and analyzing graph structure data. Due to the dataset possessing incomplete structure and insufficient labels, the ability of the GNN model to learn node features is limited. Graph contrastive learning (GCL) effectively mitigates the labels problem, however, it ignores more structural information and ignores the semantic information between subgraphs. In this paper, we propose a novel Graph Augmentation and Local-aware Multi-view Graph Contrastive Learning (GL-GCL) framework to model the neighborhood information of nodes from both global and local views. Specifically, we use the original graph as the first level and use different graph augmentation methods to obtain the global structure graph as the second level. For the each level, we approach the problem from both local and global views. For the local view, we construct semantic subgraphs using node rankings, which are not limited to the first-order neighbors. The target nodes of each subgraph are input to the shared GNN encoder to obtain local-level target node embeddings. In addition, we generate local-level subgraph embeddings for subgraphs using an average pooling function. For the global view, considering that the global structure graph with sufficient structure information, we utilize a shared GNN encoder and multilayer perceptron (MLP) to learn node embeddings at the global level. The proposed GL-GCL model maximizes the common information between similar instances at multi-scales through a multi-level contrastive loss function. Extensive experiments on six real-world benchmark datasets show that our approach outperforms state-of-the-art baselines on both node classification and link prediction tasks.