Neighbor Contrastive Learning on Learnable Graph Augmentation

Recent years, graph contrastive learning (GCL), which aims to learn representations from unlabeled graphs, has made great progress. However, the existing GCL methods mostly adopt human-designed augmentations, are sensitive various datasets. In addition, losses originally developed in computer vision have been directly applied data, where neighboring nodes regarded as negatives and consequently pushed far apart anchor. this is contradictory with homophily assumption of net-works that connected often belong same class should be close each other. work, we propose an end-to-end automatic method, named NCLA apply neighbor on learnable augmentation. Several augmented views adaptive topology automatically learned by multi-head attention mechanism, can compatible datasets without prior domain knowledge. a loss devised allow multiple positives per anchor taking network supervised signals. Both augmentations embeddings proposed NCLA. Extensive experiments benchmark demonstrate yields state-of-the-art node classification performance self-supervised even exceeds ones, when labels extremely limited. Our code released at https://github.com/shenxiaocam/NCLA.