Knowledge Graph Edits Can Displace Correct Answers in Model Rankings

Summary

Editing a knowledge graph embedding (KGE) model to promote a desired answer can push other correct answers out of the returned ranking, even when they do not reuse the edited parameter. The paper introduces a rank-displacement audit at three scopes: facts supported by the edited parameter, other correct answers for the target query, and correct answers for queries sharing the same relation. It also derives dimensional and geometric conditions under which an update can improve the target while preserving selected scores exactly. On FB15k-237 using DistMult and ComplEx, direct promotion consistently moved the target into the top ten, but avoided measurable damage in only 23.0-23.2% of edits. Strict score preservation avoided measured damage but succeeded in only 1.3-1.4% of cases. Support-regularized entity editing achieved the strongest joint success rate, 36.3-37.7%. Rank-truncated preservation reached 32.8-34.7% and reduced the mean number of displaced answers from about 14 to 1.2. Results across embedding dimensions, scoring functions, ranking conventions, and a learned editor indicate that locality depends on both the protected scope and the editing mechanism. The authors argue that KGE editing evaluations should report correction success together with the incidence and severity of rank displacement.