Graph SSL

Graph self-supervised learning studies how to learn useful node, edge, and whole-graph representations from graph structure and features when labeled examples are scarce. Common approaches include contrastive learning across augmented graph views, masked graph autoencoders, and predictive objectives that learn to predict representations across time or graph scales. These representations are evaluated on tasks such as node classification, link prediction, graph classification, dynamic-graph anomaly detection, and molecular property prediction. A key challenge is making pretraining transfer reliably: augmentations can remove meaningful structural signals, and gains on one dataset or task do not guarantee robustness across domains or graph types.