Adaptive Graph Augmentation for Graph Contrastive Learning

Graph contrastive learning emerged as a promising method for graph representation learning. The traditional graph contrastive methods utilize data augmentations for original graphs and train models during pre-training, and for different downstream tasks and different datasets, specific finetuning approaches are required. However, there exists a long-standing gap between the performance of models during pre-training and finetuning, which directly hinders such methods to learn downstream task-relevant information during pre-training. To address this issue, we propose a novel method called adaptive graph augmentation (AGA) for self-supervised graph contrastive learning. We propose a metric called average graph confusion in pre-training to predict the performance of models on downstream tasks. Sufficient experiments prove that the proposed average graph confusion aligns well with the accuracy of downstream tasks, and there is a strong correlation between them. Besides, AGA outperforms benchmark methods in the settings of semi-supervised and unsupervised learning for graph classification. Our implementation is available at: https://github.com/sugarlemons/AGA.