Trustworthy Feature Importance Avoids Unrestricted Permutations

2026-04-13

AI Summary

This study addresses the issue of extrapolation error induced by unrestricted permutation in feature importance evaluation, which can lead to distorted results. It is the first to systematically elucidate the relationship between such errors and existing methodologies. To circumvent the pitfalls of unconstrained permutation, the work proposes three novel strategies: conditional model dependence, Gaussian-transformed Knockoffs, and constrained Accumulated Local Effects (ALE) plot design. Through rigorous theoretical analysis and comprehensive experiments, the proposed framework demonstrates a substantial reduction—and in some cases complete elimination—of extrapolation error. The approach maintains theoretical soundness while significantly enhancing the reliability and credibility of feature importance assessments.

Technology Category

• Machine Learning: Feature Construction/Reformulation
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty
• Game Theory and Economic Paradigms: Imperfect Information

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs

Abstract

Feature importance methods using unrestricted permutations are flawed due to extrapolation errors; such errors appear in all non-trivial variable importance approaches. We propose three new approaches: conditional model reliance and Knockoffs with Gaussian transformation, and restricted ALE plot designs. Theoretical and numerical results show our strategies reduce/eliminate extrapolation.

Problem

• feature importance
• permutation
• extrapolation error
• variable importance
• trustworthy

Innovation

• conditional model reliance
• Knockoffs
• Gaussian transformation
• restricted ALE plots
• extrapolation error