"""Permutation importance for estimators.""" import numbers import numpy as np from ..ensemble._bagging import _generate_indices from ..metrics import check_scoring, get_scorer_names from ..metrics._scorer import _check_multimetric_scoring, _MultimetricScorer from ..model_selection._validation import _aggregate_score_dicts from ..utils import Bunch, _safe_indexing, check_array, check_random_state from ..utils._param_validation import ( HasMethods, Integral, Interval, RealNotInt, StrOptions, validate_params, ) from ..utils.parallel import Parallel, delayed def _weights_scorer(scorer, estimator, X, y, sample_weight): if sample_weight is not None: return scorer(estimator, X, y, sample_weight=sample_weight) return scorer(estimator, X, y) def _calculate_permutation_scores( estimator, X, y, sample_weight, col_idx, random_state, n_repeats, scorer, max_samples, ): """Calculate score when `col_idx` is permuted.""" random_state = check_random_state(random_state) # Work on a copy of X to ensure thread-safety in case of threading based # parallelism. Furthermore, making a copy is also useful when the joblib # backend is 'loky' (default) or the old 'multiprocessing': in those cases, # if X is large it will be automatically be backed by a readonly memory map # (memmap). X.copy() on the other hand is always guaranteed to return a # writable data-structure whose columns can be shuffled inplace. if max_samples < X.shape[0]: row_indices = _generate_indices( random_state=random_state, bootstrap=False, n_population=X.shape[0], n_samples=max_samples, ) X_permuted = _safe_indexing(X, row_indices, axis=0) y = _safe_indexing(y, row_indices, axis=0) if sample_weight is not None: sample_weight = _safe_indexing(sample_weight, row_indices, axis=0) else: X_permuted = X.copy() scores = [] shuffling_idx = np.arange(X_permuted.shape[0]) for _ in range(n_repeats): random_state.shuffle(shuffling_idx) if hasattr(X_permuted, "iloc"): col = X_permuted.iloc[shuffling_idx, col_idx] col.index = X_permuted.index X_permuted[X_permuted.columns[col_idx]] = col else: X_permuted[:, col_idx] = X_permuted[shuffling_idx, col_idx] scores.append(_weights_scorer(scorer, estimator, X_permuted, y, sample_weight)) if isinstance(scores[0], dict): scores = _aggregate_score_dicts(scores) else: scores = np.array(scores) return scores def _create_importances_bunch(baseline_score, permuted_score): """Compute the importances as the decrease in score. Parameters ---------- baseline_score : ndarray of shape (n_features,) The baseline score without permutation. permuted_score : ndarray of shape (n_features, n_repeats) The permuted scores for the `n` repetitions. Returns ------- importances : :class:`~sklearn.utils.Bunch` Dictionary-like object, with the following attributes. importances_mean : ndarray, shape (n_features, ) Mean of feature importance over `n_repeats`. importances_std : ndarray, shape (n_features, ) Standard deviation over `n_repeats`. importances : ndarray, shape (n_features, n_repeats) Raw permutation importance scores. """ importances = baseline_score - permuted_score return Bunch( importances_mean=np.mean(importances, axis=1), importances_std=np.std(importances, axis=1), importances=importances, ) @validate_params( { "estimator": [HasMethods(["fit"])], "X": ["array-like"], "y": ["array-like", None], "scoring": [ StrOptions(set(get_scorer_names())), callable, list, tuple, dict, None, ], "n_repeats": [Interval(Integral, 1, None, closed="left")], "n_jobs": [Integral, None], "random_state": ["random_state"], "sample_weight": ["array-like", None], "max_samples": [ Interval(Integral, 1, None, closed="left"), Interval(RealNotInt, 0, 1, closed="right"), ], }, prefer_skip_nested_validation=True, ) def permutation_importance( estimator, X, y, *, scoring=None, n_repeats=5, n_jobs=None, random_state=None, sample_weight=None, max_samples=1.0, ): """Permutation importance for feature evaluation [BRE]_. The :term:`estimator` is required to be a fitted estimator. `X` can be the data set used to train the estimator or a hold-out set. The permutation importance of a feature is calculated as follows. First, a baseline metric, defined by :term:`scoring`, is evaluated on a (potentially different) dataset defined by the `X`. Next, a feature column from the validation set is permuted and the metric is evaluated again. The permutation importance is defined to be the difference between the baseline metric and metric from permutating the feature column. Read more in the :ref:`User Guide `. Parameters ---------- estimator : object An estimator that has already been :term:`fitted` and is compatible with :term:`scorer`. X : ndarray or DataFrame, shape (n_samples, n_features) Data on which permutation importance will be computed. y : array-like or None, shape (n_samples, ) or (n_samples, n_classes) Targets for supervised or `None` for unsupervised. scoring : str, callable, list, tuple, or dict, default=None Scorer to use. If `scoring` represents a single score, one can use: - a single string (see :ref:`scoring_parameter`); - a callable (see :ref:`scoring`) that returns a single value. If `scoring` represents multiple scores, one can use: - a list or tuple of unique strings; - a callable returning a dictionary where the keys are the metric names and the values are the metric scores; - a dictionary with metric names as keys and callables a values. Passing multiple scores to `scoring` is more efficient than calling `permutation_importance` for each of the scores as it reuses predictions to avoid redundant computation. If None, the estimator's default scorer is used. n_repeats : int, default=5 Number of times to permute a feature. n_jobs : int or None, default=None Number of jobs to run in parallel. The computation is done by computing permutation score for each columns and parallelized over the columns. `None` means 1 unless in a :obj:`joblib.parallel_backend` context. `-1` means using all processors. See :term:`Glossary ` for more details. random_state : int, RandomState instance, default=None Pseudo-random number generator to control the permutations of each feature. Pass an int to get reproducible results across function calls. See :term:`Glossary `. sample_weight : array-like of shape (n_samples,), default=None Sample weights used in scoring. .. versionadded:: 0.24 max_samples : int or float, default=1.0 The number of samples to draw from X to compute feature importance in each repeat (without replacement). - If int, then draw `max_samples` samples. - If float, then draw `max_samples * X.shape[0]` samples. - If `max_samples` is equal to `1.0` or `X.shape[0]`, all samples will be used. While using this option may provide less accurate importance estimates, it keeps the method tractable when evaluating feature importance on large datasets. In combination with `n_repeats`, this allows to control the computational speed vs statistical accuracy trade-off of this method. .. versionadded:: 1.0 Returns ------- result : :class:`~sklearn.utils.Bunch` or dict of such instances Dictionary-like object, with the following attributes. importances_mean : ndarray of shape (n_features, ) Mean of feature importance over `n_repeats`. importances_std : ndarray of shape (n_features, ) Standard deviation over `n_repeats`. importances : ndarray of shape (n_features, n_repeats) Raw permutation importance scores. If there are multiple scoring metrics in the scoring parameter `result` is a dict with scorer names as keys (e.g. 'roc_auc') and `Bunch` objects like above as values. References ---------- .. [BRE] :doi:`L. Breiman, "Random Forests", Machine Learning, 45(1), 5-32, 2001. <10.1023/A:1010933404324>` Examples -------- >>> from sklearn.linear_model import LogisticRegression >>> from sklearn.inspection import permutation_importance >>> X = [[1, 9, 9],[1, 9, 9],[1, 9, 9], ... [0, 9, 9],[0, 9, 9],[0, 9, 9]] >>> y = [1, 1, 1, 0, 0, 0] >>> clf = LogisticRegression().fit(X, y) >>> result = permutation_importance(clf, X, y, n_repeats=10, ... random_state=0) >>> result.importances_mean array([0.4666..., 0. , 0. ]) >>> result.importances_std array([0.2211..., 0. , 0. ]) """ if not hasattr(X, "iloc"): X = check_array(X, force_all_finite="allow-nan", dtype=None) # Precompute random seed from the random state to be used # to get a fresh independent RandomState instance for each # parallel call to _calculate_permutation_scores, irrespective of # the fact that variables are shared or not depending on the active # joblib backend (sequential, thread-based or process-based). random_state = check_random_state(random_state) random_seed = random_state.randint(np.iinfo(np.int32).max + 1) if not isinstance(max_samples, numbers.Integral): max_samples = int(max_samples * X.shape[0]) elif max_samples > X.shape[0]: raise ValueError("max_samples must be <= n_samples") if callable(scoring): scorer = scoring elif scoring is None or isinstance(scoring, str): scorer = check_scoring(estimator, scoring=scoring) else: scorers_dict = _check_multimetric_scoring(estimator, scoring) scorer = _MultimetricScorer(scorers=scorers_dict) baseline_score = _weights_scorer(scorer, estimator, X, y, sample_weight) scores = Parallel(n_jobs=n_jobs)( delayed(_calculate_permutation_scores)( estimator, X, y, sample_weight, col_idx, random_seed, n_repeats, scorer, max_samples, ) for col_idx in range(X.shape[1]) ) if isinstance(baseline_score, dict): return { name: _create_importances_bunch( baseline_score[name], # unpack the permuted scores np.array([scores[col_idx][name] for col_idx in range(X.shape[1])]), ) for name in baseline_score } else: return _create_importances_bunch(baseline_score, np.array(scores))