How are feature_importances in RandomForestClassifier determined?

Abstraction: Scikit-learn random forest feature importance via weighted impurity reduction

Key points:

• Feature importance is the sum of weighted impurity decreases at all nodes where a feature is used as a split: n_samples impurity - n_left impurity_left - n_right * impurity_right
• Accumulated importances are divided by root node's weighted_n_node_samples, then optionally normalized to sum to 1.0
• Implemented in Cython (compute_feature_importances) iterating over all tree nodes
• Verified manually on iris dataset: computed values match clf.feature_importances_ (petal width=0.923, petal length=0.064)
• All classes assumed to have equal weight (weight=1); be careful about weighted cases

Connections: Scikit Learn · Random Forest · Feature Importance · Decision Trees

Source: http://stackoverflow.com/questions/15810339/how-are-feature-importances-in-randomforestclassifier-determined/15821880#15821880