A module for computing feature importances by measuring how score decreases
when a feature is not available. It contains basic building blocks;
there is a full-featured sklearn-compatible implementation
A similar method is described in Breiman, “Random Forests”, Machine Learning, 45(1), 5-32, 2001 (available online at https://www.stat.berkeley.edu/%7Ebreiman/randomforest2001.pdf), with an application to random forests. It is known in literature as “Mean Decrease Accuracy (MDA)” or “permutation importance”.
get_score_importances(score_func, X, y, n_iter=5, columns_to_shuffle=None, random_state=None)¶
(base_score, score_decreases)tuple with the base score and score decreases when a feature is not available.
score_decreasesis a list of length
n_iterwith feature importance arrays (each array is of shape
n_features); feature importances are computed as score decrease when a feature is not available.
n_iteriterations of the basic algorithm is done, each iteration starting from a different random seed.
If you just want feature importances, you can take a mean of the result:
import numpy as np from eli5.permutation_importance import get_scores_importances base_score, score_decreases = get_score_importances(score_func, X, y) feature_importances = np.mean(score_decreases, axis=0)
iter_shuffled(X, columns_to_shuffle=None, pre_shuffle=False, random_state=None)¶
Return an iterator of X matrices which have one or more columns shuffled. After each iteration yielded matrix is mutated inplace, so if you want to use multiple of them at the same time, make copies.
columns_to_shuffleis a sequence of column numbers to shuffle. By default, all columns are shuffled once, i.e. columns_to_shuffle is
pre_shuffleis True, a copy of
Xis shuffled once, and then result takes shuffled columns from this copy. If it is False, columns are shuffled on fly.
pre_shuffle = Truecan be faster if there is a lot of columns, or if columns are used multiple times.