Unifying Post-hoc Explanations of Knowledge Graph Completions

A new paper calls for a unified taxonomy to bring order to the fragmented field of post-hoc explainability for Knowledge Graph Completion (KGC), arguing that the lack of formalization and consistent evaluations is hindering reproducibility and cross-study comparisons [1][2]. Knowledge Graphs structure information as entity-relation-entity triples, and KGC is the machine-learning task of predicting plausible missing triples [1][2]. Post-hoc explainability in this domain seeks to identify which existing triples most influence a model's predictions [1][2]. The paper, submitted on 29 Jul 2025 and revised on 15 Jun 2026, proposes a characterization of post-hoc explanations through multi-objective optimization, aiming to unify existing algorithms and the explanations they produce by balancing effectiveness and conciseness [1][2]. A central contribution is the examination of improved evaluation protocols built on metrics such as Mean Reciprocal Rank and Hits@k, tested through illustrative experiments [1][2]. The authors stress that interpretability should be measured by the ability of explanations to address queries meaningful to end users, not just technical benchmarks [1][2]. This push for standardization echoes broader concerns in machine learning about reproducibility. For instance, a separate review of quantum circuit generation systems found that while all reviewed systems addressed syntactic validity, none reported end-to-end evaluation on quantum hardware, leaving a significant gap between generated artifacts and practical deployment [5]. The paper's focus on identifying which triples act as causal factors for a prediction touches on foundational concepts. In philosophy, causality is understood as an influence where one event contributes to the production of another, with the cause being at least partly responsible for the effect [3]. The paper's multi-objective framework formalizes this search for influential triples within a KGC model's reasoning. The work is hosted on arXiv, a platform deeply integrated with community tools: Hugging Face's Paper Pages, for example, automatically extract arXiv IDs to link papers with models, datasets, and interactive demos, and even allow authors to claim verified authorship [6][7]. The paper's call for unified methods and evaluation standards aims to make KGC explainability research more reproducible and impactful [1][2]. By proposing a formal taxonomy and scrutinizing evaluation protocols, the work addresses a structural weakness in a field where, as the authors note, inconsistent practices have made cross-study comparisons difficult [1][2].