[python-package] feature_names_in_ spaces replaced by underlines incompatible with sklearn

python-package/lightgbm/basic.py

@@ -1781,6 +1781,7 @@ def __init__(
         #
         # This is here mostly for scikit-learn's benefit, as it tracks whether input data had feature names.
         self._has_non_default_feature_names: bool = False
+        self.original_feature_name_: np.ndarray = np.array([], dtype=object)
 
     def __del__(self) -> None:
         try:
@@ -2061,14 +2062,19 @@ def _lazy_init(
                 categorical_feature=categorical_feature,
                 pandas_categorical=self.pandas_categorical,
             )
+            self.original_feature_name_ = np.array(feature_name, dtype=object)
         if nwd.is_into_dataframe(data) and feature_name == "auto":
             feature_name = nw.from_native(data).schema.names()
+            self.original_feature_name_ = np.array(feature_name, dtype=object)
 
         # 'feature_name == "auto"' after the block above means no feature names were provided
         # by either the data type (DataFrame/pyarrow) or the user's 'feature_name' argument.
         # LightGBM will assign auto-generated names like Column_0, Column_1, etc.
         self._has_non_default_feature_names = feature_name != "auto"
 
+        if feature_name != "auto":
+            self.original_feature_name_ = np.array(feature_name, dtype=object)
+
         # process for args
         params = {} if params is None else params
         args_names = inspect.signature(self.__class__._lazy_init).parameters.keys()
@@ -2966,6 +2972,7 @@ def set_feature_name(self, feature_name: _LGBM_FeatureNameConfiguration) -> "Dat
         if feature_name != "auto":
             self.feature_name = feature_name
             self._has_non_default_feature_names = True
+            self.original_feature_name_ = np.array(feature_name, dtype=object)
         if self._handle is not None and feature_name is not None and feature_name != "auto":
             if len(feature_name) != self.num_feature():
                 raise ValueError(

python-package/lightgbm/sklearn.py

@@ -716,6 +716,7 @@ def __init__(
         self._classes: Optional[np.ndarray] = None
         self._n_classes: int = -1
         self.set_params(**kwargs)
+        self.original_feature_name_: np.ndarray = np.array([], dtype=object)
 
     # scikit-learn 1.6 introduced an __sklearn__tags() method intended to replace _more_tags().
     # _more_tags() can be removed whenever lightgbm's minimum supported scikit-learn version
@@ -1146,6 +1147,8 @@ def fit(
         # is set BEFORE fitting.
         self._n_features = self._Booster.num_feature()
 
+        self.original_feature_name_ = train_set.original_feature_name_
+
         # This attribute informs self.features_in_, but isn't set until here
         # because Dataset.construct(), called by train(), is responsible for updating it.
         self._fitted_with_feature_names = train_set._has_non_default_feature_names
@@ -1382,7 +1385,7 @@ def feature_names_in_(self) -> np.ndarray:
                 "The training data did not have feature names "
                 "(e.g. was a numpy array rather than a pandas DataFrame)."
             )
-        return np.array(self.feature_name_)
+        return np.array(self.original_feature_name_, dtype=object)
 
     @feature_names_in_.deleter
     def feature_names_in_(self) -> None:

tests/python_package_test/test_sklearn.py

@@ -1801,7 +1801,7 @@ def test_feature_names_in_and_predict_warning(
         # feature_name_: always accessible, reflects actual names used internally
         # feature_names_in_: absent when no named features, present otherwise
         if fit_X_type in types_with_feat_names:
-            np_assert_array_equal(model.feature_names_in_, np.array(col_names), strict=True)
+            np_assert_array_equal(model.feature_names_in_, np.array(col_names, dtype=object), strict=True)
             assert model.feature_name_ == col_names
         else:
             assert model.feature_name_ == default_names
@@ -1832,9 +1832,9 @@ def test_feature_names_in_and_predict_warning(
     model = lgb.LGBMClassifier(n_estimators=2, num_leaves=3).fit(X_fit, y, feature_name=custom_names)
 
     # feature names from keyword arg should be used, not any from the input data
-    np_assert_array_equal(model.feature_names_in_, np.array(custom_names), strict=True)
+    np_assert_array_equal(model.feature_names_in_, np.array(custom_names, dtype=object), strict=True)
     assert model.feature_name_ == custom_names
-    np_assert_array_equal(model.feature_names_in_, np.array(custom_names), strict=True)
+    np_assert_array_equal(model.feature_names_in_, np.array(custom_names, dtype=object), strict=True)
     assert model.n_features_in_ == n_features
 
     # predict() should not raise a warning if input has feature names
@@ -1872,6 +1872,32 @@ def _get_expected_failed_tests(estimator):
     return estimator._more_tags()["_xfail_checks"]
 
 
+def test_feature_names_in_():
+    """
+    Test that feature_names_in_ returns the same feature names as the input.
+    """
+    pd = pytest.importorskip("pandas")
+
+    X = pd.DataFrame(
+        {
+            "feature age": [1.0, 2.0, 3.0, 4.0],
+            "body mass index": [5.0, 6.0, 7.0, 8.0],
+        }
+    )
+    y = np.array([0, 1, 0, 1])
+
+    model = lgb.LGBMClassifier(
+        n_estimators=2,
+        num_leaves=3,
+        verbosity=-1,
+    ).fit(X, y)
+
+    np_assert_array_equal(
+        model.feature_names_in_,
+        X.columns.to_numpy(dtype=object),
+        strict=True,
+    )
+
 @parametrize_with_checks(
     [ExtendedLGBMClassifier(), ExtendedLGBMRegressor(), lgb.LGBMClassifier(), lgb.LGBMRegressor()],
     expected_failed_checks=_get_expected_failed_tests,