Permutation based feature importance

This code estimate the importances of the features of a fastai tabular learner model using the permutation method.

While I made some modifications to it (adding functionalities and fixing bugs) most of the credits goes to Miguel Mota Pinto for the first prototype, Zachary Mueller for the improved version and John Keefe for the plotting function.

Usage

from feature_importance import *

# gets feature importances as a dataframe
importances = feature_importance(learn)

# note that you can, optionnally, specify the dataset on which you want to compute the importances (validation dataset by default)
importances_train = feature_importance(learn, DatasetType.Train)

# plots feature importances directly
plot_feature_importance(learn)

Todo

• the code severely needs refactoring to improve readability
• no need to sort features when computing feature importance (better done only if the user needs it for plotting purposes)
• could output a class that inherit from a dataframe but with an overloaded plotting function

For more unofficial fastai extensions, see the Fastai Extensions Repository.