From eb67d473b95c9709c45876c9c8bf0252fe2d1021 Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Mon, 22 Sep 2025 10:03:13 +0200
Subject: [PATCH 01/10] base logic and test

---
 sklearn/feature_selection/_base.py          | 19 ++++++++--
 sklearn/feature_selection/_rfe.py           |  1 +
 sklearn/feature_selection/tests/test_rfe.py | 40 +++++++++++++++++++--
 3 files changed, 56 insertions(+), 4 deletions(-)

diff --git a/sklearn/feature_selection/_base.py b/sklearn/feature_selection/_base.py
index 3c12cd035d5c8..02dcc2e25abef 100644
--- a/sklearn/feature_selection/_base.py
+++ b/sklearn/feature_selection/_base.py
@@ -3,6 +3,7 @@
 # Authors: The scikit-learn developers
 # SPDX-License-Identifier: BSD-3-Clause
 
+import inspect
 import warnings
 from abc import ABCMeta, abstractmethod
 from operator import attrgetter
@@ -196,7 +197,9 @@ def get_feature_names_out(self, input_features=None):
         return input_features[self.get_support()]
 
 
-def _get_feature_importances(estimator, getter, transform_func=None, norm_order=1):
+def _get_feature_importances(
+    estimator, getter, feature_indices=None, transform_func=None, norm_order=1
+):
     """
     Retrieve and aggregate (ndim > 1)  the feature importances
     from an estimator. Also optionally applies transformation.
@@ -215,6 +218,14 @@ def _get_feature_importances(estimator, getter, transform_func=None, norm_order=
         The transform to apply to the feature importances. By default (`None`)
         no transformation is applied.
 
+    feature_indices : ndarray of shape (n_features,), default=None
+        The indices of features from the full dataset whose importance are currently
+        evaluated. These are passed to `getter` when it can accept them which allows
+        using RFE with permutation importance, as shown in this documentation example:
+
+        #TODO: create and link here a documentation example showing how to use
+        # `permutation_importance` with RFE(CV).
+
     norm_order : int, default=1
         The norm order to apply when `transform_func="norm"`. Only applied
         when `importances.ndim > 1`.
@@ -243,7 +254,11 @@ def _get_feature_importances(estimator, getter, transform_func=None, norm_order=
     elif not callable(getter):
         raise ValueError("`importance_getter` has to be a string or `callable`")
 
-    importances = getter(estimator)
+    param_names = list(inspect.signature(getter).parameters.keys())
+    if "feature_indices" in param_names:
+        importances = getter(estimator, feature_indices)
+    else:
+        importances = getter(estimator)
 
     if transform_func is None:
         return importances
diff --git a/sklearn/feature_selection/_rfe.py b/sklearn/feature_selection/_rfe.py
index 056bb0203b187..07e18664b1c44 100644
--- a/sklearn/feature_selection/_rfe.py
+++ b/sklearn/feature_selection/_rfe.py
@@ -346,6 +346,7 @@ def _fit(self, X, y, step_score=None, **fit_params):
             importances = _get_feature_importances(
                 estimator,
                 self.importance_getter,
+                features,
                 transform_func="square",
             )
             ranks = np.argsort(importances)
diff --git a/sklearn/feature_selection/tests/test_rfe.py b/sklearn/feature_selection/tests/test_rfe.py
index 1f5672545874c..c707cda566705 100644
--- a/sklearn/feature_selection/tests/test_rfe.py
+++ b/sklearn/feature_selection/tests/test_rfe.py
@@ -14,12 +14,13 @@
 from sklearn.compose import TransformedTargetRegressor
 from sklearn.cross_decomposition import CCA, PLSCanonical, PLSRegression
 from sklearn.datasets import load_iris, make_classification, make_friedman1
-from sklearn.ensemble import RandomForestClassifier
+from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor
 from sklearn.feature_selection import RFE, RFECV
 from sklearn.impute import SimpleImputer
+from sklearn.inspection import permutation_importance
 from sklearn.linear_model import LinearRegression, LogisticRegression
 from sklearn.metrics import get_scorer, make_scorer, zero_one_loss
-from sklearn.model_selection import GroupKFold, cross_val_score
+from sklearn.model_selection import GroupKFold, cross_val_score, train_test_split
 from sklearn.pipeline import make_pipeline
 from sklearn.preprocessing import StandardScaler
 from sklearn.svm import SVC, SVR, LinearSVR
@@ -753,3 +754,38 @@ def test_results_per_cv_in_rfecv(global_random_seed):
     assert len(rfecv.cv_results_["split1_ranking"]) == len(
         rfecv.cv_results_["split2_ranking"]
     )
+
+
+def test_rfe_with_permutation_importance(global_random_seed):
+    """
+    Ensure that using permutation_importance as a importance_getter returns the right
+    amount of features.
+    """
+    X, y = make_friedman1(n_samples=100, n_features=7, random_state=global_random_seed)
+
+    X_train, X_test, y_train, y_test = train_test_split(
+        X, y, test_size=0.5, random_state=global_random_seed
+    )
+
+    reg = RandomForestRegressor(random_state=global_random_seed, n_estimators=2)
+
+    def permutation_importance_getter(
+        model, feature_indices, X_test, y_test, global_random_seed
+    ):
+        return permutation_importance(
+            model,
+            X_test[:, feature_indices],
+            y_test,
+            n_repeats=2,
+            random_state=global_random_seed,
+        ).importances_mean
+
+    rfe = RFE(
+        estimator=reg,
+        importance_getter=lambda model, feature_indices: permutation_importance_getter(
+            model, feature_indices, X_test, y_test, global_random_seed
+        ),
+        n_features_to_select=5,
+    ).fit(X_train, y_train)
+
+    assert rfe.n_features_ == 5

From 55db7fb528c0ce104c88f7bb9f3ce5c76aac45ec Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Mon, 22 Sep 2025 16:51:46 +0200
Subject: [PATCH 02/10] dont inspect signature for attrgetter to avoid
 python3.10 bug

---
 sklearn/feature_selection/_base.py | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/sklearn/feature_selection/_base.py b/sklearn/feature_selection/_base.py
index 02dcc2e25abef..444c7d4be5d0c 100644
--- a/sklearn/feature_selection/_base.py
+++ b/sklearn/feature_selection/_base.py
@@ -254,11 +254,14 @@ def _get_feature_importances(
     elif not callable(getter):
         raise ValueError("`importance_getter` has to be a string or `callable`")
 
-    param_names = list(inspect.signature(getter).parameters.keys())
-    if "feature_indices" in param_names:
-        importances = getter(estimator, feature_indices)
-    else:
+    if isinstance(getter, attrgetter):
         importances = getter(estimator)
+    else:
+        param_names = list(inspect.signature(getter).parameters.keys())
+        if "feature_indices" in param_names:
+            importances = getter(estimator, feature_indices)
+        else:
+            importances = getter(estimator)
 
     if transform_func is None:
         return importances

From e9a58bcf0c5f3729840029d0c00eab090e33f2b7 Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Thu, 25 Sep 2025 10:21:56 +0200
Subject: [PATCH 03/10] add changelog

---
 .../sklearn.feature_selection/32251.feature.rst          | 9 +++++++++
 1 file changed, 9 insertions(+)
 create mode 100644 doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst

diff --git a/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
new file mode 100644
index 0000000000000..903b759a94816
--- /dev/null
+++ b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
@@ -0,0 +1,9 @@
+- :class:`feature_selection.RFE` and :class:`feature_selection.RFECV`
+  now support the use of :func:`permutation_importance` as an :attr:`importance_getter`.
+  When a callable, and when it can accept it, the :attr:`importance_getter` will be passed
+  :attr:`feature_indices` along the fitted estimator.
+  The attribute :attr:`feature_indices` stores the index of the features from the full dataset
+  that have not been eliminated yet.
+  This allows methods like :func:`permutation_importance` to extract the relevant features
+  from its test set.
+  By :user:`Gaétan de Castellane <GaetandeCast>`.

From 5777a4ad0ff5f7b10b3447dd786c09f13baaa73f Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Thu, 25 Sep 2025 10:36:31 +0200
Subject: [PATCH 04/10] add test from code review

---
 sklearn/feature_selection/tests/test_rfe.py | 41 ++++++++++++++++++++-
 1 file changed, 39 insertions(+), 2 deletions(-)

diff --git a/sklearn/feature_selection/tests/test_rfe.py b/sklearn/feature_selection/tests/test_rfe.py
index c707cda566705..2ea0e258ac52d 100644
--- a/sklearn/feature_selection/tests/test_rfe.py
+++ b/sklearn/feature_selection/tests/test_rfe.py
@@ -758,8 +758,8 @@ def test_results_per_cv_in_rfecv(global_random_seed):
 
 def test_rfe_with_permutation_importance(global_random_seed):
     """
-    Ensure that using permutation_importance as a importance_getter returns the right
-    amount of features.
+    Ensure that using permutation_importance as a importance_getter returns the amount
+    features set with `n_features_to_select`.
     """
     X, y = make_friedman1(n_samples=100, n_features=7, random_state=global_random_seed)
 
@@ -789,3 +789,40 @@ def permutation_importance_getter(
     ).fit(X_train, y_train)
 
     assert rfe.n_features_ == 5
+
+
+def test_rfecv_with_permutation_importance(global_random_seed):
+    """
+    Ensure that using permutation_importance as a importance_getter in `RFECV` returns
+    exactly the five informative features of `make_friedman1`.
+    """
+    X, y = make_friedman1(
+        n_samples=1_000, n_features=7, random_state=global_random_seed
+    )
+
+    X_train, X_test, y_train, y_test = train_test_split(
+        X, y, test_size=0.5, random_state=global_random_seed
+    )
+
+    reg = RandomForestRegressor(random_state=global_random_seed, n_estimators=15)
+
+    def permutation_importance_getter(
+        model, feature_indices, X_test, y_test, global_random_seed
+    ):
+        return permutation_importance(
+            model,
+            X_test[:, feature_indices],
+            y_test,
+            n_repeats=2,
+            random_state=global_random_seed,
+        ).importances_mean
+
+    rfe = RFECV(
+        estimator=reg,
+        importance_getter=lambda model, feature_indices: permutation_importance_getter(
+            model, feature_indices, X_test, y_test, global_random_seed
+        ),
+    ).fit(X_train, y_train)
+
+    assert rfe.n_features_ == 5
+    assert_array_equal(rfe.support_, np.array(([True] * 5) + ([False] * 2)))

From f2bf623e3ebcdb012bdc005d052d46376a5c41b0 Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Thu, 25 Sep 2025 11:58:28 +0200
Subject: [PATCH 05/10] doc example and docstrings

---
 .../32251.feature.rst                         |  8 +-
 .../plot_rfe_with_cross_validation.py         | 75 +++++++++++++++++++
 sklearn/feature_selection/_base.py            |  6 +-
 sklearn/feature_selection/_rfe.py             | 16 +++-
 sklearn/feature_selection/tests/test_rfe.py   | 14 ++--
 5 files changed, 101 insertions(+), 18 deletions(-)

diff --git a/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
index 903b759a94816..c156fafc75bd1 100644
--- a/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
+++ b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
@@ -1,9 +1,9 @@
 - :class:`feature_selection.RFE` and :class:`feature_selection.RFECV`
   now support the use of :func:`permutation_importance` as an :attr:`importance_getter`.
-  When a callable, and when it can accept it, the :attr:`importance_getter` will be passed
-  :attr:`feature_indices` along the fitted estimator.
+  When a callable, and when it can accept it, the :attr:`importance_getter` is passed
+  :attr:`feature_indices` along with the fitted estimator.
   The attribute :attr:`feature_indices` stores the index of the features from the full dataset
   that have not been eliminated yet.
-  This allows methods like :func:`permutation_importance` to extract the relevant features
-  from its test set.
+  This allows methods that need a test set, like :func:`permutation_importance`, to know which
+  features of to use in their predictions.
   By :user:`Gaétan de Castellane <GaetandeCast>`.
diff --git a/examples/feature_selection/plot_rfe_with_cross_validation.py b/examples/feature_selection/plot_rfe_with_cross_validation.py
index 307707c5aa069..5f39167547518 100644
--- a/examples/feature_selection/plot_rfe_with_cross_validation.py
+++ b/examples/feature_selection/plot_rfe_with_cross_validation.py
@@ -113,3 +113,78 @@
 # In the five folds, the selected features are consistent. This is good news,
 # it means that the selection is stable across folds, and it confirms that
 # these features are the most informative ones.
+
+# %%
+# Using `permutation_importance` to select features
+# -------------------------------------------------
+# Under the hood, `RFECV` uses importance scores derived from the coefficients of the
+# linear model we used to choose which feature to eliminate. We show here how to use
+# `permutation_importance` as an alternative way to measure the importance of features.
+# For that, we need to feed the `importance_getter` parameter of RFECV a callable
+# that accepts a fitted model and an array containing the indices of the features that
+# have not been eliminated yet.
+
+# %%
+from sklearn.inspection import permutation_importance
+
+# Permutation importance need test data to produce reliable importance measures.
+X_test, y_test = make_classification(
+    n_samples=500,
+    n_features=n_features,
+    n_informative=3,
+    n_redundant=2,
+    n_repeated=0,
+    n_classes=8,
+    n_clusters_per_class=1,
+    class_sep=0.8,
+    random_state=0,
+)
+
+
+# Use `feature_indices` to extract the features that have not been eliminated yet from
+# test set.
+def permutation_importance_getter(model, feature_indices, X_test, y_test, random_state):
+    return permutation_importance(
+        model,
+        X_test[:, feature_indices],
+        y_test,
+        random_state=random_state,
+    ).importances_mean
+
+
+rfecv = RFECV(
+    estimator=clf,
+    step=1,
+    cv=cv,
+    scoring="accuracy",
+    min_features_to_select=min_features_to_select,
+    n_jobs=2,
+    importance_getter=lambda model, feature_indices: permutation_importance_getter(
+        model, feature_indices, X_test, y_test, random_state=0
+    ),
+)
+rfecv.fit(X, y)
+
+print(f"Optimal number of features: {rfecv.n_features_}")
+
+# %%
+data = {
+    key: value
+    for key, value in rfecv.cv_results_.items()
+    if key in ["n_features", "mean_test_score", "std_test_score"]
+}
+cv_results = pd.DataFrame(data)
+plt.figure()
+plt.xlabel("Number of features selected")
+plt.ylabel("Mean test accuracy")
+plt.errorbar(
+    x=cv_results["n_features"],
+    y=cv_results["mean_test_score"],
+    yerr=cv_results["std_test_score"],
+)
+plt.title("Recursive Feature Elimination \nwith correlated features")
+plt.show()
+
+# %%
+# We see that we obtain very similar results with this model agnostic feature importance
+# method.
diff --git a/sklearn/feature_selection/_base.py b/sklearn/feature_selection/_base.py
index 444c7d4be5d0c..63eeaf532af4d 100644
--- a/sklearn/feature_selection/_base.py
+++ b/sklearn/feature_selection/_base.py
@@ -220,11 +220,7 @@ def _get_feature_importances(
 
     feature_indices : ndarray of shape (n_features,), default=None
         The indices of features from the full dataset whose importance are currently
-        evaluated. These are passed to `getter` when it can accept them which allows
-        using RFE with permutation importance, as shown in this documentation example:
-
-        #TODO: create and link here a documentation example showing how to use
-        # `permutation_importance` with RFE(CV).
+        evaluated. These are passed to `getter` when it can accept them.
 
     norm_order : int, default=1
         The norm order to apply when `transform_func="norm"`. Only applied
diff --git a/sklearn/feature_selection/_rfe.py b/sklearn/feature_selection/_rfe.py
index 07e18664b1c44..f74e3495649a7 100644
--- a/sklearn/feature_selection/_rfe.py
+++ b/sklearn/feature_selection/_rfe.py
@@ -124,7 +124,13 @@ class RFE(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
 
         If `callable`, overrides the default feature importance getter.
         The callable is passed with the fitted estimator and it should
-        return importance for each feature.
+        return importance for each feature. When  it accepts it, the callable is passed
+        `feature_indices` which stores the index of the features in the full dataset
+        that have not been eliminated yet.
+
+        `feature_indices` allows RFE to be used with permutation importance, as
+        shown on RFECV at the end of
+        :ref:`sphx_glr_auto_examples_feature_selection_plot_rfe_with_cross_validation.py`.
 
         .. versionadded:: 0.24
 
@@ -647,7 +653,13 @@ class RFECV(RFE):
 
         If `callable`, overrides the default feature importance getter.
         The callable is passed with the fitted estimator and it should
-        return importance for each feature.
+        return importance for each feature. When it accepts it, the callable is passed
+        `feature_indices` which stores the index of the features in the full dataset
+        that have not been eliminated yet.
+
+        `feature_indices` allows RFECV to be used with permutation importance, as
+        shown at the end of
+        :ref:`sphx_glr_auto_examples_feature_selection_plot_rfe_with_cross_validation.py`.
 
         .. versionadded:: 0.24
 
diff --git a/sklearn/feature_selection/tests/test_rfe.py b/sklearn/feature_selection/tests/test_rfe.py
index 2ea0e258ac52d..d5a5dd3cf9c99 100644
--- a/sklearn/feature_selection/tests/test_rfe.py
+++ b/sklearn/feature_selection/tests/test_rfe.py
@@ -770,14 +770,14 @@ def test_rfe_with_permutation_importance(global_random_seed):
     reg = RandomForestRegressor(random_state=global_random_seed, n_estimators=2)
 
     def permutation_importance_getter(
-        model, feature_indices, X_test, y_test, global_random_seed
+        model, feature_indices, X_test, y_test, random_state
     ):
         return permutation_importance(
             model,
             X_test[:, feature_indices],
             y_test,
             n_repeats=2,
-            random_state=global_random_seed,
+            random_state=random_state,
         ).importances_mean
 
     rfe = RFE(
@@ -807,22 +807,22 @@ def test_rfecv_with_permutation_importance(global_random_seed):
     reg = RandomForestRegressor(random_state=global_random_seed, n_estimators=15)
 
     def permutation_importance_getter(
-        model, feature_indices, X_test, y_test, global_random_seed
+        model, feature_indices, X_test, y_test, random_state
     ):
         return permutation_importance(
             model,
             X_test[:, feature_indices],
             y_test,
             n_repeats=2,
-            random_state=global_random_seed,
+            random_state=random_state,
         ).importances_mean
 
-    rfe = RFECV(
+    rfecv = RFECV(
         estimator=reg,
         importance_getter=lambda model, feature_indices: permutation_importance_getter(
             model, feature_indices, X_test, y_test, global_random_seed
         ),
     ).fit(X_train, y_train)
 
-    assert rfe.n_features_ == 5
-    assert_array_equal(rfe.support_, np.array(([True] * 5) + ([False] * 2)))
+    assert rfecv.n_features_ == 5
+    assert_array_equal(rfecv.support_, np.array(([True] * 5) + ([False] * 2)))

From 701e6a6ec6442935f3c92671f2d908ca66470b5c Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Thu, 25 Sep 2025 15:01:59 +0200
Subject: [PATCH 06/10] typo and reformulation

---
 .../plot_rfe_with_cross_validation.py              | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/examples/feature_selection/plot_rfe_with_cross_validation.py b/examples/feature_selection/plot_rfe_with_cross_validation.py
index 5f39167547518..0b1eff3ee464d 100644
--- a/examples/feature_selection/plot_rfe_with_cross_validation.py
+++ b/examples/feature_selection/plot_rfe_with_cross_validation.py
@@ -118,16 +118,16 @@
 # Using `permutation_importance` to select features
 # -------------------------------------------------
 # Under the hood, `RFECV` uses importance scores derived from the coefficients of the
-# linear model we used to choose which feature to eliminate. We show here how to use
+# linear model we used, to choose which feature to eliminate. We show here how to use
 # `permutation_importance` as an alternative way to measure the importance of features.
-# For that, we need to feed the `importance_getter` parameter of RFECV a callable
-# that accepts a fitted model and an array containing the indices of the features that
-# have not been eliminated yet.
+# For that, we use a callable in the `importance_getter` parameter of RFECV.
+# This callable accepts a fitted model and an array containing the indices of the
+# features that have not been eliminated yet.
 
 # %%
 from sklearn.inspection import permutation_importance
 
-# Permutation importance need test data to produce reliable importance measures.
+# Permutation importance needs test data to produce reliable importance measures.
 X_test, y_test = make_classification(
     n_samples=500,
     n_features=n_features,
@@ -141,8 +141,8 @@
 )
 
 
-# Use `feature_indices` to extract the features that have not been eliminated yet from
-# test set.
+# Use `feature_indices` to extract from the test set the features that have not been
+# eliminated yet.
 def permutation_importance_getter(model, feature_indices, X_test, y_test, random_state):
     return permutation_importance(
         model,

From a7897d32038028c09761a1b8611e3b86ca4dd998 Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Thu, 25 Sep 2025 18:33:24 +0200
Subject: [PATCH 07/10] add missing backtick and colon

---
 sklearn/feature_selection/_rfe.py | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/sklearn/feature_selection/_rfe.py b/sklearn/feature_selection/_rfe.py
index f74e3495649a7..a1aa088c4651a 100644
--- a/sklearn/feature_selection/_rfe.py
+++ b/sklearn/feature_selection/_rfe.py
@@ -120,7 +120,7 @@ class RFE(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
         For example, give `regressor_.coef_` in case of
         :class:`~sklearn.compose.TransformedTargetRegressor`  or
         `named_steps.clf.feature_importances_` in case of
-        class:`~sklearn.pipeline.Pipeline` with its last step named `clf`.
+        :class:`~sklearn.pipeline.Pipeline` with its last step named `clf`.
 
         If `callable`, overrides the default feature importance getter.
         The callable is passed with the fitted estimator and it should
@@ -128,8 +128,8 @@ class RFE(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
         `feature_indices` which stores the index of the features in the full dataset
         that have not been eliminated yet.
 
-        `feature_indices` allows RFE to be used with permutation importance, as
-        shown on RFECV at the end of
+        `feature_indices` allows `RFE` to be used with permutation importance, as
+        shown on `RFECV` at the end of
         :ref:`sphx_glr_auto_examples_feature_selection_plot_rfe_with_cross_validation.py`.
 
         .. versionadded:: 0.24
@@ -657,7 +657,7 @@ class RFECV(RFE):
         `feature_indices` which stores the index of the features in the full dataset
         that have not been eliminated yet.
 
-        `feature_indices` allows RFECV to be used with permutation importance, as
+        `feature_indices` allows `RFECV` to be used with permutation importance, as
         shown at the end of
         :ref:`sphx_glr_auto_examples_feature_selection_plot_rfe_with_cross_validation.py`.
 

From db66df65daef9aca83c1ccc469670a9c1e09414d Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Fri, 26 Sep 2025 09:42:42 +0200
Subject: [PATCH 08/10] typo

---
 .../sklearn.feature_selection/32251.feature.rst                 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
index c156fafc75bd1..509ed1a81d1e7 100644
--- a/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
+++ b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
@@ -5,5 +5,5 @@
   The attribute :attr:`feature_indices` stores the index of the features from the full dataset
   that have not been eliminated yet.
   This allows methods that need a test set, like :func:`permutation_importance`, to know which
-  features of to use in their predictions.
+  features to use in their predictions.
   By :user:`Gaétan de Castellane <GaetandeCast>`.

From 38549548e1332c8b9942d5d53788ce9832f8720c Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Fri, 26 Sep 2025 15:54:07 +0200
Subject: [PATCH 09/10] improve docstring and fix example

---
 .../plot_rfe_with_cross_validation.py          |  4 +++-
 sklearn/feature_selection/_rfe.py              | 18 +++++++++++++-----
 2 files changed, 16 insertions(+), 6 deletions(-)

diff --git a/examples/feature_selection/plot_rfe_with_cross_validation.py b/examples/feature_selection/plot_rfe_with_cross_validation.py
index 0b1eff3ee464d..78c38caf78c01 100644
--- a/examples/feature_selection/plot_rfe_with_cross_validation.py
+++ b/examples/feature_selection/plot_rfe_with_cross_validation.py
@@ -137,7 +137,7 @@
     n_classes=8,
     n_clusters_per_class=1,
     class_sep=0.8,
-    random_state=0,
+    random_state=1,
 )
 
 
@@ -148,6 +148,8 @@ def permutation_importance_getter(model, feature_indices, X_test, y_test, random
         model,
         X_test[:, feature_indices],
         y_test,
+        n_repeats=10,
+        n_jobs=2,
         random_state=random_state,
     ).importances_mean
 
diff --git a/sklearn/feature_selection/_rfe.py b/sklearn/feature_selection/_rfe.py
index a1aa088c4651a..cbd9a745ec8b1 100644
--- a/sklearn/feature_selection/_rfe.py
+++ b/sklearn/feature_selection/_rfe.py
@@ -124,9 +124,9 @@ class RFE(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
 
         If `callable`, overrides the default feature importance getter.
         The callable is passed with the fitted estimator and it should
-        return importance for each feature. When  it accepts it, the callable is passed
+        return importance for each feature. When it accepts it, the callable is passed
         `feature_indices` which stores the index of the features in the full dataset
-        that have not been eliminated yet.
+        that have not yet been eliminated in previous iterations.
 
         `feature_indices` allows `RFE` to be used with permutation importance, as
         shown on `RFECV` at the end of
@@ -134,6 +134,10 @@ class RFE(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
 
         .. versionadded:: 0.24
 
+        .. versionchanged:: 1.8
+            Add support for passing `feature_indices` to the callable when part of its
+            signature.
+
     Attributes
     ----------
     classes_ : ndarray of shape (n_classes,)
@@ -655,14 +659,18 @@ class RFECV(RFE):
         The callable is passed with the fitted estimator and it should
         return importance for each feature. When it accepts it, the callable is passed
         `feature_indices` which stores the index of the features in the full dataset
-        that have not been eliminated yet.
+        that have not yet been eliminated in previous iterations.
 
-        `feature_indices` allows `RFECV` to be used with permutation importance, as
-        shown at the end of
+        `feature_indices` allows `RFE` to be used with permutation importance, as
+        shown on `RFECV` at the end of
         :ref:`sphx_glr_auto_examples_feature_selection_plot_rfe_with_cross_validation.py`.
 
         .. versionadded:: 0.24
 
+        .. versionchanged:: 1.8
+            Add support for passing `feature_indices` to the callable when part of its
+            signature.
+
     Attributes
     ----------
     classes_ : ndarray of shape (n_classes,)

From a6ed7216f9b52b2dc0e85d851a1050536ef7a93e Mon Sep 17 00:00:00 2001
From: Gaetan <gaetan@probabl.ai>
Date: Fri, 26 Sep 2025 17:49:33 +0200
Subject: [PATCH 10/10] improve documentation

---
 .../32251.feature.rst                         |  2 +-
 .../plot_rfe_with_cross_validation.py         | 33 ++++++++-----------
 sklearn/feature_selection/_rfe.py             | 12 +++----
 3 files changed, 20 insertions(+), 27 deletions(-)

diff --git a/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
index 509ed1a81d1e7..7fb6e8806ac51 100644
--- a/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
+++ b/doc/whats_new/upcoming_changes/sklearn.feature_selection/32251.feature.rst
@@ -1,6 +1,6 @@
 - :class:`feature_selection.RFE` and :class:`feature_selection.RFECV`
   now support the use of :func:`permutation_importance` as an :attr:`importance_getter`.
-  When a callable, and when it can accept it, the :attr:`importance_getter` is passed
+  When a callable, and when possible, the :attr:`importance_getter` is passed
   :attr:`feature_indices` along with the fitted estimator.
   The attribute :attr:`feature_indices` stores the index of the features from the full dataset
   that have not been eliminated yet.
diff --git a/examples/feature_selection/plot_rfe_with_cross_validation.py b/examples/feature_selection/plot_rfe_with_cross_validation.py
index 78c38caf78c01..13e92598a0116 100644
--- a/examples/feature_selection/plot_rfe_with_cross_validation.py
+++ b/examples/feature_selection/plot_rfe_with_cross_validation.py
@@ -21,12 +21,13 @@
 # features are non-informative as they are drawn at random.
 
 from sklearn.datasets import make_classification
+from sklearn.model_selection import train_test_split
 
 n_features = 15
 feat_names = [f"feature_{i}" for i in range(15)]
 
 X, y = make_classification(
-    n_samples=500,
+    n_samples=1_000,
     n_features=n_features,
     n_informative=3,
     n_redundant=2,
@@ -36,6 +37,9 @@
     class_sep=0.8,
     random_state=0,
 )
+X_train, X_test, y_train, y_test = train_test_split(
+    X, y, train_size=0.5, shuffle=False, random_state=0
+)
 
 # %%
 # Model training and selection
@@ -60,7 +64,7 @@
     min_features_to_select=min_features_to_select,
     n_jobs=2,
 )
-rfecv.fit(X, y)
+rfecv.fit(X_train, y_train)
 
 print(f"Optimal number of features: {rfecv.n_features_}")
 
@@ -117,29 +121,18 @@
 # %%
 # Using `permutation_importance` to select features
 # -------------------------------------------------
-# Under the hood, `RFECV` uses importance scores derived from the coefficients of the
-# linear model we used, to choose which feature to eliminate. We show here how to use
-# `permutation_importance` as an alternative way to measure the importance of features.
-# For that, we use a callable in the `importance_getter` parameter of RFECV.
+# The `importance_getter` parameter in RFE and RFECV uses by default the `coef_` (e.g.
+# in linear models) or the `feature_importances_` attributes of an estimator to derive
+# feature importance. These importance measures are used to choose which features to
+# eliminate first.
+#
+# We show here how to use a callable to compute the `permutation_importance` instead.
 # This callable accepts a fitted model and an array containing the indices of the
-# features that have not been eliminated yet.
+# features that remain after elimination.
 
 # %%
 from sklearn.inspection import permutation_importance
 
-# Permutation importance needs test data to produce reliable importance measures.
-X_test, y_test = make_classification(
-    n_samples=500,
-    n_features=n_features,
-    n_informative=3,
-    n_redundant=2,
-    n_repeated=0,
-    n_classes=8,
-    n_clusters_per_class=1,
-    class_sep=0.8,
-    random_state=1,
-)
-
 
 # Use `feature_indices` to extract from the test set the features that have not been
 # eliminated yet.
diff --git a/sklearn/feature_selection/_rfe.py b/sklearn/feature_selection/_rfe.py
index cbd9a745ec8b1..053ff76e67840 100644
--- a/sklearn/feature_selection/_rfe.py
+++ b/sklearn/feature_selection/_rfe.py
@@ -124,9 +124,9 @@ class RFE(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
 
         If `callable`, overrides the default feature importance getter.
         The callable is passed with the fitted estimator and it should
-        return importance for each feature. When it accepts it, the callable is passed
-        `feature_indices` which stores the index of the features in the full dataset
-        that have not yet been eliminated in previous iterations.
+        return importance for each feature. When the callable also accepts
+        `feature_indices` in its signature, it will be passed the index of the features
+        of the full dataset that remain after elimination in previous iterations.
 
         `feature_indices` allows `RFE` to be used with permutation importance, as
         shown on `RFECV` at the end of
@@ -657,9 +657,9 @@ class RFECV(RFE):
 
         If `callable`, overrides the default feature importance getter.
         The callable is passed with the fitted estimator and it should
-        return importance for each feature. When it accepts it, the callable is passed
-        `feature_indices` which stores the index of the features in the full dataset
-        that have not yet been eliminated in previous iterations.
+        return importance for each feature. When the callable also accepts
+        `feature_indices` in its signature, it will be passed the index of the features
+        of the full dataset that remain after elimination in previous iterations.
 
         `feature_indices` allows `RFE` to be used with permutation importance, as
         shown on `RFECV` at the end of