GCL-ALG: graph contrastive learning with adaptive learnable view generators

Data augmentation is a pivotal part of graph contrastive learning, which can mine implicit graph data information to improve the quality of representation learning. Research on graph data augmentation has achieved promising results in recent years. However, existing graph contrastive learning methods are trapped in inherent predefined augmentation schemes, which greatly limits the generalization of augmentation methods. To this end, we propose a new adaptive original topology learnable data augmentation algorithm, Graph Contrastive Learning with Adaptive Learnable View Generators (GCL-ALG), to optimize the augmentation process and feature learning in an end-to-end self-supervised learning approach. Specifically, GCL-ALG introduces graph neural networks (GNN), graph attention modules and edge probability distributions to build a dual-level feature extraction framework to generate highly reliable representations, while integrating network science theory to selectively modify the strength of augmentation probabilities from node-level and edge-level, and then train dynamically learnable augmentation instances. Moreover, GCL-ALG designs multiple loss functions to drive the representation optimization to ensure that the generated graph representations are highly discriminative across different tasks. Extensive experiments are conducted on unsupervised learning, semi-supervised learning and transfer learning application tasks. The experimental results demonstrate the superior performance of the proposed GCL-ALG method on 16 benchmark datasets.