When we train a machine learning model, we often ask: Which features matter most? For linear models, this question seems straightforward: we inspect coefficients. But for nonlinear models, ensembles, or neural networks, this becomes less obvious. Permutation Feature Importance offers a simple and powerful idea: if a feature is important, then disrupting it should harm predictive performance. The method is conceptually minimal:
- Train a model normally.
- Measure its performance on a validation set.
- Randomly permute the values of one feature.
- Measure performance again.
The idea is simple:
- If performance drops significantly, the feature was informative.
- If performance remains unchanged, the feature was not contributing.
