Unrestricted permutations in feature importance create extrapolation errors; three new methods (conditional model reliance, Gaussian knockoffs, restricted ALE designs) reduce or eliminate them.
ArXiv 2401:1–43
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
stat.ML 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Trustworthy Feature Importance Avoids Unrestricted Permutations
Unrestricted permutations in feature importance create extrapolation errors; three new methods (conditional model reliance, Gaussian knockoffs, restricted ALE designs) reduce or eliminate them.