When Successful Knowledge Graph Edits Displace Correct Answers: A Rank-Displacement Audit

Forum topic · 小凯 · 2026-09-15

Summary

Editing a knowledge graph embedding (KGE) model to promote a desired answer can inadvertently push other correct answers out of the returned list. Standard locality tests that only check facts reusing the edited parameter may miss these ranking side effects. This paper introduces a rank-displacement audit across three scopes: facts supported by the edited parameter, other correct answers to the target query, and correct answers in queries sharing the same relation. The authors derive dimensional and geometric conditions under which an update improves the target while exactly preserving selected scores. On FB15k-237 with DistMult and ComplEx, direct promotion always moves the target into the top ten, but only 23.0-23.2% of edits are damage-free; strict preservation causes no measured damage but succeeds in only 1.3-1.4%; entity editing with supporting regularization achieves the best joint success rate of 36.3-37.7%; rank-truncation preservation reaches 32.8-34.7% and cuts the average number of displaced answers from about 14 to 1.2. The work recommends that KGE edits report both correction success and rank-displacement severity.

Paper Overview

Field: Machine Learning
Authors: Yi-Cheng Lai, Jerry Wang, Hsin-Ling Hsu, Li-Chu Chi, Ya-Wen Teng, Hen-Hsen Huang
Published: 2026-09-15
arXiv: 2609.12116

Abstract

Editing a knowledge graph embedding (KGE) model to promote a desired answer can displace correct answers from the returned list. Locality tests based only on facts that reuse the edited parameter can miss this ranking effect. The authors introduce a common rank-displacement audit at three scopes: facts supported by the edited parameter, other correct answers to the target query, and correct answers across queries with the same relation. They also derive dimensional and geometric conditions for an update to improve the target while exactly preserving selected scores.

Key Findings

Experiments on FB15k-237 with DistMult and ComplEx show:

Direct promotion always moves the target into the top ten, but is damage-free in only 23.0–23.2% of edits.
Strict preservation causes no measured damage, yet succeeds in only 1.3–1.4% of cases.
Entity editing with supporting regularization achieves the highest joint success rate of 36.3–37.7%.
Truncated-rank preservation reaches 32.8–34.7% and reduces the average number of displaced answers from roughly 14 to 1.2.

Implications

Experiments across dimensions, scorers, ranking conventions, and a learned editor show that locality depends jointly on the protected scope and the editing mechanism. The authors conclude that KGE editing evaluations should report both the correction success rate and the incidence and severity of rank displacement.