Measures how much model performance decreases when the feature is permuted.
- It is model-agnostic and particularly useful for non-linear models.
Features that are deemed of low importance for a bad model could be important for a good model.
The difference between Permutation feature importance and Gini importance is that Gini importance is determined during training based on how much each feature contributes to internal representation decisions in trees, while Permutation Importance measures performance degradation post-hoc in black box models (model-agnostic).