Graph Contrastive Learning

Graph contrastive learning uses self-supervision to teach models which nodes, subgraphs, or whole graphs should have similar representations, reducing reliance on labeled data. Common approaches create augmented graph views and contrast their embeddings, using graph neural networks, spectral or structural signals, and objectives at multiple scales; some methods instead focus on selecting reliable positives and negatives. The learned representations support tasks such as node classification, clustering, link prediction, recommendation, and anomaly detection, including on dynamic and heterogeneous graphs. A central challenge is designing augmentations and contrastive pairs that remove noise without discarding useful structure, especially when graphs are sparse, heterophilous, or changing over time.