A Graph Contrastive Learning Framework with Adaptive Augmentation and Encoding for Unaligned Views

Abstract

Recently, graph contrastive learning has emerged as a successful method for graph representation learning, but it still faces three challenging problems. First, existing contrastive methods cannot preserve the semantics of the graph well after view augmentation. Second, most models use the same encoding method to encode homophilic and heterophilic graphs, failing to obtain better-quality representations. Finally, most models require that the two augmented views have the same set of nodes, which limits flexible augmentation methods. To address the above problems, we propose a novel graph contrastive learning framework with adaptive augmentation and encoding for unaligned views, called GCAUV in this paper. First, we propose multiple node centrality metrics to compute edge centrality for view augmentation, adaptively removing edges with low centrality to preserve the semantics of the graph well. Second, we use a multi-headed graph attention network to encode homophilic graphs, and use MLP to encode heterophilic graphs. Finally, we propose g-EMD distance instead of cosine similarity to measure the distance between positive and negative samples. We also perform adversarial training by adding perturbation to node features to improve the accuracy of GCAUV. Experimental results show that our method outperforms the state-of-the-art graph contrastive methods on node classification tasks.