4 / 6 Interpretability

Why can permutation importance underrate correlated features?

A model relies on either of two strongly correlated variables interchangeably. Why might ordinary one-feature-at-a-time permutation importance assign both variables low importance?

A

Correlated variables necessarily receive identical permutation-importance values

B

Permutation importance is calculated only from tree split counts

C

Permutation preserves every conditional dependency involving the shuffled variable

D

When one variable is shuffled, its correlated substitute can preserve much of the model's predictive performance