featureimpact

featureimpact let's you compute the statistical impact of features given a trained estimator. The computation is based on the mean variation of the difference between perturbed and original predictions. The estimator must predict purely numerical values. All features must also consist of purely numerical values.

Example:

from featureimpact import FeatureImpact
fi = FeatureImpact()
fi.make_quantiles(X_train)
impact = fi.compute_impact(model, X_test)

Note: In order to run the examples you'll need scikit-learn installed in addition to this package and its regular dependencies.

The algorithm is described here: https://bloomen.github.io/pub/featureimpact.pdf