FB15k-237 — a tensor to complete, not compress

A knowledge graph is a set of facts (subject, relation, object). FB15k-237’s training split is a binary tensor of 14,505 × 237 × 14,505: fifty billion cells, a quarter of a million ones, a density of about 5 × 10⁻⁶. Toutanova and Chen (2015) cut it from FB15k by removing near-duplicate and inverse relations.

The job is completion: score facts nobody recorded. Unobserved is not false — nobody recorded most true facts — so models train with a ranking or classification loss over observed facts, not least squares over every cell.

Relations split cleanly. Most, carrying about four in five training triples, are directional: their reversal is never recorded. Across them only about 3% of pairs are exposed (the object is also a subject of that relation, or the subject an object). The exception is a hierarchy: in /location/location/contains, four in five pairs are exposed.