During pre-training, the GNN is trained on a link prediction task. By sampling nodes and forming triplets, a pre-training loss is constructed that improves the similarity between the contextual subgraphs of two candidate nodes.

→ View original post on X — @maximelabonne,