Graph neural networks can combine shared-neighbour structure with node and edge features for link prediction.

Traditional predictors include common neighbours, Adamic–Adar, preferential attachment and Katz index. Graph neural networks instead learn representations by aggregating information from graph structure and node or edge attributes, with node-based and subgraph-based approaches offering different expressive power. This can improve on a single local heuristic when relevant features exist, but incomplete graphs, hidden confounding and distribution shift still limit interpretation and out-of-sample reliability.