Link Prediction

Link prediction asks which connections are missing or likely to form, supporting tasks such as knowledge-graph completion, recommendations, collaboration discovery, and forecasting interactions over time. Methods range from neighborhood heuristics and graph embeddings to GNNs, subgraph Transformers, temporal models, and language- or multimodal approaches that combine graph structure with text or other features. Evaluation spans static and temporal benchmarks, but results can depend heavily on negative sampling, data splits, and simple persistence or count-based baselines. Key challenges include generalizing to unseen entities or graphs, scaling to large networks, and balancing predictive accuracy with fairness, stability, and reliable evaluation.