Knowledge graphs have a blind spot, and a new paper argues the field has been looking in the wrong place to fix it.
Most knowledge graph completion research focuses on triplet prediction — essentially asking whether a specific subject-relation-object link is plausible. A paper published on arXiv introduces a parallel task called relation set completion (RSC), which asks a different question: given an entity and its known relations, which semantically compatible relations are missing entirely? The authors also propose RelSetE, a model that learns latent patterns among an entity's observed relations to infer the absent ones. Three benchmark datasets derived from standard knowledge graph benchmarks are included, and code and data are publicly available.
The distinction matters because triplet-level incompleteness and entity-relation compatibility incompleteness are separate problems. A system that confidently scores individual links can still miss whole categories of relations that should apply to an entity — a gap that quietly degrades downstream applications like search, recommendation, and question answering that rely on these graphs.
Knowledge graph completion has been a crowded research area for years, with models like TransE, RotatE, and their descendants stacking incremental gains on link prediction benchmarks. Framing RSC as a complementary task rather than a replacement is the pragmatic move — it sidesteps the argument that link prediction is obsolete while still pointing at a real limitation the field has glossed over.