In self-supervised learning with graph representations, a model is trained to learn useful features of nodes or entire graphs without explicit labels.
Typical methods use tasks such as predicting neighboring nodes, reconstructing subgraphs, or distinguishing between real and corrupted connections.
Graph Neural Networks (GNNs) are often used to generate rich node or graph embeddings by aggregating neighborhood information.
The goal is to obtain structured and informative representations that can be used for various applications such as classification, clustering, or link prediction.