Self-supervised learning with graphs is an approach where Graph Neural Networks (GNNs) are used to learn meaningful representations from unlabeled graph data.
Special tasks are constructed, such as predicting missing edges, reconstructing node attributes, or comparing subgraphs, to train the model.
These so-called pretext tasks allow the model to capture structural and semantic information of the graph without relying on labeled data.
Typical methods include contrastive learning techniques, where positive and negative sample examples are generated to learn the similarity of nodes or graphs.
Such self-supervised graph models improve generalization on various tasks like node classification, link prediction, or graph classification.