COIN: Graph Contrastive Learning with Orthogonal Continuous Augmentation and Information Balance

Graph contrastive learning has demonstrated success in generating high-quality node representations. Existing data augmentation and pretext tasks for unsupervised graph contrastive learning reslut in suboptimal node representations. In this paper, we propose COIN, a new graph contrastive learning framework with continuous orthogonal augmentation and information balancing, to improve model performance. We introduce a continuous view augmenter with a masked topology module and a cross-channel feature module to adaptively capture both topological and feature information, preventing dimensional collapse. We further propose the InfoBal principle to balance consistency and diversity across views, and to enable effective utilization of representational information. Extensive experiments demonstrate that COIN outperforms state-of-the-art frameworks.