Local-Global Fusion Augmented Graph Contrastive Learning Based on Generative Models

Abstract

Recently, graph contrastive learning (GCL) has emerged as a dominant technique for unsupervised graph representation learning. Existing GCL-based approaches typically adopt data augmentation to generate two contrastive views, and maximize the similarity between representations that share the same semantics in these two views. However, there are two weaknesses in existing methods: 1) They are highly dependent on the selection of augmentation modes, such as random node and edge perturbation as well as node feature masking, which may lead to changes in the structure and semantic information of the graph. 2) Existing augmentation strategies fail to adequately model local and global information. Graph-based generative models can use graph topology and semantic information to reconstruct node features, so as to avoid random perturbation of features. Therefore, we design a novel graph contrastive learning approach based on generative model augmentation that minimizes damage to the original graph and incorporates local-global information into the augmented view. Specifically, we first obtain the diffusion graph with global information by diffusion method. Based on the structural and semantic information of the original graph and the diffused graph, local and global augmentations are performed using generative models respectively. Then, we fuse the local and global information into the final augmented view. Extensive experiments illustrate the superior performance of model over state-of-the-art methods.