AFANS: Augmentation-Free Graph Contrastive Learning with Adversarial Negative Sampling

Graph Contrastive Learning (GCL) has emerged as a highly promising methodology in graph representation learning, mainly due to its label-independent nature. The construction of positive and negative samples is crucial for the effectiveness of GCL. Yet, the accurate identification of genuine positive and negative samples poses a formidable challenge. Traditional methods employ data augmentation for constructing positive samples, which confronts two significant impediments: i) the potential distortion of semantic integrity and ii) the difficulty in devising augmentation strategies universally applicable across diverse datasets. In the realm of negative sample generation, reliance on in-batch negative samples or a memory bank may engender “easy” or “false” negative samples, detrimentally impacting model performance. To address these issues, we propose to use Exponential Moving Average (EMA) instead of data augmentation to construct effective positive samples. Additionally, we introduce an adversarial generator to fabricate more challenging negative samples, which are subsequently amalgamated with in-batch negative samples. Furthermore, we implement two constrained loss functions aimed at reducing redundancy amongst negative samples, enriching the model with more salient information. The effectiveness of our proposed method is validated through an unsupervised graph classification task on five real-world datasets. Empirical results substantiate that our approach surpasses current state-of-the-art methodologies.