ENH Add clip parameter to MaxAbsScaler (scikit-learn#31790)

Co-authored-by: Jérémie du Boisberranger <jeremie@probabl.ai>

Author: glevv

doc/whats_new/upcoming_changes/sklearn.preprocessing/31790.enhancement.rst

@@ -0,0 +1,3 @@
+- :class:`preprocessing.MaxAbsScaler` can now clip out-of-range values in held-out data
+  with the parameter `clip`.
+  By :user:`Hleb Levitski <glevv>`.

sklearn/preprocessing/_data.py

@@ -329,7 +329,16 @@ class MinMaxScaler(OneToOneFeatureMixin, TransformerMixin, BaseEstimator):
 
     clip : bool, default=False
         Set to True to clip transformed values of held-out data to
-        provided `feature range`.
+        provided `feature_range`.
+        Since this parameter will clip values, `inverse_transform` may not
+        be able to restore the original data.
+
+        .. note::
+            Setting `clip=True` does not prevent feature drift (a distribution
+            shift between training and test data). The transformed values are clipped
+            to the `feature_range`, which helps avoid unintended behavior in models
+            sensitive to out-of-range inputs (e.g. linear models). Use with care,
+            as clipping can distort the distribution of test data.
 
         .. versionadded:: 0.24
 
@@ -1172,6 +1181,18 @@ class MaxAbsScaler(OneToOneFeatureMixin, TransformerMixin, BaseEstimator):
         Set to False to perform inplace scaling and avoid a copy (if the input
         is already a numpy array).
 
+    clip : bool, default=False
+        Set to True to clip transformed values of held-out data to [-1, 1].
+        Since this parameter will clip values, `inverse_transform` may not
+        be able to restore the original data.
+
+        .. note::
+            Setting `clip=True` does not prevent feature drift (a distribution
+            shift between training and test data). The transformed values are clipped
+            to the [-1, 1] range, which helps avoid unintended behavior in models
+            sensitive to out-of-range inputs (e.g. linear models). Use with care,
+            as clipping can distort the distribution of test data.
+
     Attributes
     ----------
     scale_ : ndarray of shape (n_features,)
@@ -1222,10 +1243,14 @@ class MaxAbsScaler(OneToOneFeatureMixin, TransformerMixin, BaseEstimator):
            [ 0. ,  1. , -0.5]])
     """
 
-    _parameter_constraints: dict = {"copy": ["boolean"]}
+    _parameter_constraints: dict = {
+        "copy": ["boolean"],
+        "clip": ["boolean"],
+    }
 
-    def __init__(self, *, copy=True):
+    def __init__(self, *, copy=True, clip=False):
         self.copy = copy
+        self.clip = clip
 
     def _reset(self):
         """Reset internal data-dependent state of the scaler, if necessary.
@@ -1340,8 +1365,20 @@ def transform(self, X):
 
         if sparse.issparse(X):
             inplace_column_scale(X, 1.0 / self.scale_)
+            if self.clip:
+                np.clip(X.data, -1.0, 1.0, out=X.data)
         else:
             X /= self.scale_
+            if self.clip:
+                device_ = device(X)
+                X = _modify_in_place_if_numpy(
+                    xp,
+                    xp.clip,
+                    X,
+                    xp.asarray(-1.0, dtype=X.dtype, device=device_),
+                    xp.asarray(1.0, dtype=X.dtype, device=device_),
+                    out=X,
+                )
         return X
 
     def inverse_transform(self, X):

sklearn/preprocessing/tests/test_common.py

@@ -42,7 +42,7 @@ def _get_valid_samples_by_column(X, col):
 @pytest.mark.parametrize(
     "est, func, support_sparse, strictly_positive, omit_kwargs",
     [
-        (MaxAbsScaler(), maxabs_scale, True, False, []),
+        (MaxAbsScaler(), maxabs_scale, True, False, ["clip"]),
         (MinMaxScaler(), minmax_scale, False, False, ["clip"]),
         (StandardScaler(), scale, False, False, []),
         (StandardScaler(with_mean=False), scale, True, False, []),

sklearn/preprocessing/tests/test_data.py

@@ -707,6 +707,7 @@ def test_standard_check_array_of_inverse_transform():
     "estimator",
     [
         MaxAbsScaler(),
+        MaxAbsScaler(clip=True),
         MinMaxScaler(),
         MinMaxScaler(clip=True),
         KernelCenterer(),
@@ -2517,6 +2518,8 @@ def test_minmax_scaler_clip(feature_range):
     # test behaviour of the parameter 'clip' in MinMaxScaler
     X = iris.data
     scaler = MinMaxScaler(feature_range=feature_range, clip=True).fit(X)
+    # create a test sample with features outside the training feature range:
+    # first 2 features < min(X) and last 2 features > max(X)
     X_min, X_max = np.min(X, axis=0), np.max(X, axis=0)
     X_test = [np.r_[X_min[:2] - 10, X_max[2:] + 10]]
     X_transformed = scaler.transform(X_test)
@@ -2526,6 +2529,25 @@ def test_minmax_scaler_clip(feature_range):
     )
 
 
+@pytest.mark.parametrize(
+    "data_constructor", [np.array] + CSC_CONTAINERS + CSR_CONTAINERS
+)
+def test_maxabs_scaler_clip(data_constructor):
+    # test behaviour of the parameter 'clip' in MaxAbsScaler
+    X = data_constructor(iris.data)
+    is_sparse = sparse.issparse(X)
+    scaler = MaxAbsScaler(clip=True).fit(X)
+    # create a test sample with features outside the training max abs range:
+    # first 2 features > max(abs(X)) and last 2 features < -max(abs(X))
+    max_abs = np.max(np.abs(X), axis=0)
+    max_abs = max_abs.data if is_sparse else max_abs
+    X_test = data_constructor(
+        np.hstack((max_abs[:2] + 10, -max_abs[2:] - 10)).reshape(1, -1)
+    )
+    X_transformed = scaler.transform(X_test)
+    assert_allclose_dense_sparse(X_transformed, data_constructor([[1, 1, -1, -1]]))
+
+
 def test_standard_scaler_raise_error_for_1d_input():
     """Check that `inverse_transform` from `StandardScaler` raises an error
     with 1D array.