Can near-zero permutation importance from a weak model prove that a feature contains no useful information?
Classic MLLevel: IntermediateImportance is model-relative
Can near-zero permutation importance from a weak model prove that a feature contains no useful information?
Hints
- The estimator is fixed during permutation.
- Unused information may still exist.
Answer
No. It measures this fitted model's reliance under the chosen evaluation data and score. A better model may exploit information that the weak model failed to learn.
Source and licence: scikit-learn 1.7.2 — permutation importance · BSD-3-Clause · answer: Quantoros editorial