permutation_importance.py

# import function, or patch it:
  # Note: may need to install mlxtend
  
try:
    from sklearn.inspection import permutation_importance
except ImportError:
    print("Problem importing permutation_importance -- patching")
    from mlxtend.evaluate import feature_importance_permutation
    def permutation_importance(estimator, X, y, scoring='r2', n_repeats=5):
        """
        Use mlxtend function, but give the same interface as the upcoming
        function merged into the 0.22.dev branch of scikit-learn.
        """
        # match the arguments for the new function to the mlxtend function
        means, values = feature_importance_permutation(X.values, y, estimator.predict, 
                                                       scoring, num_rounds=n_repeats)
        return {
            'importances': values,
            'importances_mean': means
        }


# define and fit classifier:
clf = RandomForestClassifier()
clf.fit(train_X, train_y)

# perform permutation importance
perm_importance = permutation_importance(clf, train_X, train_y, scoring='accuracy')
mean_perm_imp = perm_imp['importances_mean']
mean_pi_scaled = mean_perm_imp / mean_perm_imp.sum()

# compile & plot results
feat_imp_df = pd.DataFrame({
    'features': train_X.columns,
    'model_importances': rf_best.feature_importances_,
    'permuted_importances': mean_pi_scaled
})
feat_imp_df.sort_values(by='permuted_importances').plot.barh(x='features')