Classical ML, validation, metrics, interpretation, debugging · insight-1152
Report permutation importance with its repeat spread and the metric that matches the decision
- Evidence type
- documented fact
- Difficulty
- easy
- Roles
- Data scientist, ML engineer
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A permutation importance table shows one number per feature from a single shuffle, scored with the default metric.
Action
Use many repeats, report the mean with its standard deviation, treat features within a couple of standard deviations of zero as unproven, and compute importance under the metric the business uses.
Reason
Each repeat is a random draw, and importance is the drop in the chosen score, so both randomness and the scorer shape the result.
Boundary
Driver reports and feature pruning from a fitted model scored on held-out data.
Symptom or risk
Single shuffles are noisy, so small differences are ranked as if real, and a different metric, particularly on imbalanced data, can reorder features.
Exceptions
With a very large validation set and a stable metric, spread is small but should still be shown.
Trade-offs
More repeats and several scorers cost extra model predictions.
Pitfalls
Computing it for a weak model: the docs warn low importance for a bad model can be high for a good one.
Lessons and concepts
Claims and their sources
permutation_importance repeats shuffles (n_repeats) to give a measure of variance, depends on the scoring argument, and for imbalanced classification the choice of metric can be critical; held-out data shows which features support generalization.