Trustworthy Feature Importance Avoids Unrestricted Permutations

arXiv stat.ML / 4/14/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

Key Points

  • Existing feature-importance methods that rely on unrestricted permutations are flawed because they introduce extrapolation errors that can affect essentially all non-trivial variable importance approaches.
  • The authors argue that extrapolation failures are a common underlying problem, not an edge case limited to a specific method.
  • They propose three new strategies—(1) conditional model reliance, (2) Knockoffs with Gaussian transformation, and (3) restricted ALE plot designs—to address the extrapolation issue.
  • The paper reports theoretical and numerical results indicating these strategies can reduce or eliminate extrapolation errors, improving trustworthiness of feature importance estimates.

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.