Classical ML, validation, metrics, interpretation, debugging · insight-1153
Permute within strata when a feature is largely determined by a grouping variable
- Evidence type
- empirical finding
- Difficulty
- hard
- Roles
- Data scientist
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
Permutation importance ranks route or season features high, and one of those features is almost fully set by the airport or month column.
Action
Shuffle the feature only within bins of the variable that determines it (conditional permutation), or permute the correlated group together, and compare with the plain version.
Reason
Conditional permutation respects the dependence between predictors, and the unconditional scheme was shown to give correlated predictors an artificial advantage.
Boundary
Random forests and boosted trees with correlated or hierarchically related predictors.
Symptom or risk
Plain shuffling pairs a value with an impossible partner, so the model is scored on unrealistic rows and the importance mixes true reliance with the artifact.
Exceptions
With roughly independent features the plain version is adequate and cheaper.
Trade-offs
Strata need enough rows; many sparse strata make the shuffle nearly a no-op.
Pitfalls
Conditioning on a variable that is itself a consequence of the feature of interest, which hides its real effect.
Lessons and concepts
Claims and their sources
Strobl et al. report that random forest importance measures are biased towards correlated predictors, partly through the unconditional permutation scheme, and propose a conditional permutation scheme that reflects each predictor's impact more reliably.