Contrast Sample Pre-selection for Augmentation-Free Graph Contrastive Learning

In recent years, contrastive learning has become a dominant approach in graph self-supervised learning. However, most existing graph contrastive learning methods rely on the homophily assumption, leading to poor performance on heterophilic graphs. Developing contrastive learning methods that are effective for both homophilic and heterophilic graphs remains an important challenge. Among these methods, augmentation-free contrastive learning has gained attention due to its simplicity. However, existing approaches typically select positive samples based on similarity in the representation space, making them heavily dependent on the quality of node representations produced by the graph encoder in the early training stages. In this paper, we propose a contrastive learning method that directly determines positive sample pairs from raw graph data, avoiding the reliance on unstable early-stage sample selection. We leverage feature similarity and neighborhood similarity in the original graph data to select positive samples. Through different model parameter initialization strategies, we demonstrate that our method mitigates the adverse effects of low-quality node representations during training. Extensive experiments on various datasets further validate the effectiveness and generalizability of our method across both homophilic and heterophilic graphs.