Trustworthy Feature Importance Avoids Unrestricted Permutations
For practitioners using feature importance in interpretable ML, this work addresses a fundamental flaw in permutation-based methods, offering more reliable alternatives.
The paper identifies that unrestricted permutations in feature importance methods cause extrapolation errors, and proposes three new approaches (conditional model reliance, Knockoffs with Gaussian transformation, and restricted ALE plots) that theoretically and numerically reduce or eliminate these errors.
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.