Module eval 

Link prediction evaluation with filtered ranking.

Computes MRR and Hits@k in the filtered setting: for each test triple (h, r, t), rank all entities as head replacements and tail replacements, excluding other known-true triples from the ranking.

“Filtered” means that when computing the rank of a correct entity, we remove all other entities that also form known-true triples from the candidate set. This prevents penalizing a model for ranking another correct answer above the target. See Bordes et al. (2013) for the original protocol.

Tie-breaking

This implementation uses pessimistic tie-breaking: an entity with the same score as the target does not increase the rank. Only entities with strictly better (lower) scores count. This is equivalent to “best-case rank among tied entities.”

Parallelism

Evaluation is parallelized across test triples via rayon. Each test triple is scored independently against all entities.

Structs

EvalResult

Evaluation results with optional per-relation breakdown.

Metrics

Link prediction metrics.

Functions

evaluate_link_prediction

Evaluate link prediction on test triples in the filtered setting.

evaluate_link_prediction_detailed

Evaluate link prediction with per-relation breakdown.