Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper identifies extrapolation errors as a fundamental flaw in feature importance methods that rely on unrestricted permutations. To address this, they introduce three novel approaches: conditional model reliance, Knockoffs with Gaussian transformation, and restricted ALE plot designs. Both theoretical analysis and empirical results demonstrate that these strategies effectively mitigate or eliminate extrapolation errors, leading to more trustworthy feature importance estimates.
Feature importance estimates are likely garbage if they rely on unrestricted permutations, but these three new methods offer a fix.
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.