Change StandardScaler to the SquashScaler for the tabular_pipline

CHANGES.rst

@@ -14,6 +14,10 @@ New features
 
 Changes
 -------
+- The :func: `tabular_pipeline` uses a :class:`SquashingScaler` instead of a
+  :class:`StandardScaler` for centering and scaling numerical features
+  when linear models are used.
+  :pr:`1644` by :user:`Simon Dierickx <dierickxsimon>`
 
 Bugfixes
 --------

doc/modules/default_wrangling/tabular_pipeline.rst

@@ -6,7 +6,7 @@
 .. |HistGradientBoostingRegressor| replace:: :class:`~sklearn.ensemble.HistGradientBoostingRegressor`
 .. |HistGradientBoostingClassifier| replace:: :class:`~sklearn.ensemble.HistGradientBoostingClassifier`
 .. |Pipeline| replace:: :class:`~sklearn.pipeline.Pipeline`
-.. |StandardScaler| replace:: :class:`~sklearn.preprocessing.StandardScaler`
+.. |SquashingScaler| replace:: :class:`~sklearn.preprocessing.SquashingScaler`
 .. |SimpleImputer| replace:: :class:`~sklearn.impute.SimpleImputer`
 
 .. _user_guide_tabular_pipeline:
@@ -18,15 +18,15 @@ The |tabular_pipeline| is a function that, given a scikit-learn estimator,
 returns a full scikit-learn |Pipeline| that contains a |TableVectorizer|
 followed by the given estimator.
 If the estimator is a linear model (e.g., ``Ridge``, ``LogisticRegression``),
-|tabular_pipeline| adds a |StandardScaler| and a |SimpleImputer| to the pipeline.
+|tabular_pipeline| adds a |SquashingScaler| and a |SimpleImputer| to the pipeline.
 
 >>> from sklearn.linear_model import LinearRegression
 >>> from skrub import tabular_pipeline
 >>> tabular_pipeline(LinearRegression()) # doctest: +SKLEARN_VERSION >= "1.4" +ELLIPSIS
 Pipeline(steps=[('tablevectorizer',
                  TableVectorizer(datetime=DatetimeEncoder(periodic_encoding='spline'))),
                 ('simpleimputer', SimpleImputer(add_indicator=True)),
-                ('standardscaler', StandardScaler()),
+                ('squashingscaler', SquashingScaler(max_absolute_value=5)),
                 ('linearregression', LinearRegression())])
 
 It is also possible to call the function with the name of the task that must be
@@ -63,7 +63,7 @@ problems, but may not beat properly tuned ad-hoc pipelines.
    * - Numeric preprocessor
      - No processing
      - No processing
-     - :class:`~sklearn.preprocessing.StandardScaler`
+     - :class:`~sklearn.preprocessing.SquashingScaler`
    * - Date preprocessor
      - :class:`DatetimeEncoder`
      - :class:`DatetimeEncoder`

skrub/_tabular_pipeline.py

@@ -5,11 +5,12 @@
 from sklearn.base import BaseEstimator
 from sklearn.impute import SimpleImputer
 from sklearn.pipeline import make_pipeline
-from sklearn.preprocessing import OrdinalEncoder, StandardScaler
+from sklearn.preprocessing import OrdinalEncoder
 from sklearn.utils.fixes import parse_version
 
 from ._datetime_encoder import DatetimeEncoder
 from ._sklearn_compat import get_tags
+from ._squashing_scaler import SquashingScaler
 from ._string_encoder import StringEncoder
 from ._table_vectorizer import TableVectorizer
 from ._to_categorical import ToCategorical
@@ -99,6 +100,11 @@ def tabular_pipeline(estimator, *, n_jobs=None):
         The high cardinality encoder has been changed from
         :class:`~skrub.MinHashEncoder` to :class:`~skrub.StringEncoder`.
 
+    .. versionchanged:: 0.7.0
+        The :class:`~skrub.SquashingScaler` with `max_absolute_value=5` is now used instead of
+        :class:`~sklearn.preprocessing.StandardScaler` for centering and scaling
+        numerical features when using linear models.
+
     Parameters
     ----------
     estimator : {"regressor", "regression", "classifier", "classification"} or sklearn.base.BaseEstimator
@@ -136,7 +142,7 @@ def tabular_pipeline(estimator, *, n_jobs=None):
     - An optional :obj:`~sklearn.impute.SimpleImputer` imputes missing values by their
       mean and adds binary columns that indicate which values were missing. This step is
       only added if the ``estimator`` cannot handle missing values itself.
-    - An optional :obj:`~sklearn.preprocessing.StandardScaler` centers and rescales the
+    - An optional :obj:`~skrub.SquashingScaler` centers and rescales the
       data. This step is not added (because it is unnecessary) when the ``estimator`` is
       a tree ensemble such as random forest or gradient boosting.
     - The last step is the provided ``estimator``.
@@ -215,7 +221,7 @@ def tabular_pipeline(estimator, *, n_jobs=None):
     Pipeline(steps=[('tablevectorizer',
                     TableVectorizer(datetime=DatetimeEncoder(periodic_encoding='spline'))),
                     ('simpleimputer', SimpleImputer(add_indicator=True)),
-                    ('standardscaler', StandardScaler()),
+                    ('squashingscaler', SquashingScaler(max_absolute_value=5)),
                     ('logisticregression', LogisticRegression())])
 
     By applying only the first pipeline step we can see the transformed data that is
@@ -235,7 +241,7 @@ def tabular_pipeline(estimator, *, n_jobs=None):
     Pipeline(steps=[('tablevectorizer',
                     TableVectorizer(datetime=DatetimeEncoder(periodic_encoding='spline'))),
                     ('simpleimputer', SimpleImputer(add_indicator=True)),
-                    ('standardscaler', StandardScaler()),
+                    ('squashingscaler', SquashingScaler(max_absolute_value=5)),
                     ('logisticregression', LogisticRegression())])
 
     For a :obj:`~sklearn.linear_model.LogisticRegression`, we get:
@@ -247,7 +253,7 @@ def tabular_pipeline(estimator, *, n_jobs=None):
     - A :obj:`~sklearn.impute.SimpleImputer`, as the
       :obj:`~sklearn.linear_model.LogisticRegression` cannot handle missing values.
 
-    - A :obj:`~sklearn.preprocessing.StandardScaler` for centering and standard scaling
+    - A :obj:`~skrub.SquashingScaler` for centering and scaling
       numerical features.
 
     On the other hand, For the :obj:`~sklearn.ensemble.HistGradientBoostingClassifier`
@@ -331,6 +337,6 @@ def tabular_pipeline(estimator, *, n_jobs=None):
     if not get_tags(estimator).input_tags.allow_nan:
         steps.append(SimpleImputer(add_indicator=True))
     if not isinstance(estimator, _TREE_ENSEMBLE_CLASSES):
-        steps.append(StandardScaler())
+        steps.append(SquashingScaler(max_absolute_value=5))
     steps.append(estimator)
     return make_pipeline(*steps)

skrub/tests/test_tabular_learner.py

@@ -3,10 +3,11 @@
 from sklearn import ensemble
 from sklearn.impute import SimpleImputer
 from sklearn.linear_model import Ridge
-from sklearn.preprocessing import OneHotEncoder, OrdinalEncoder, StandardScaler
+from sklearn.preprocessing import OneHotEncoder, OrdinalEncoder
 from sklearn.utils.fixes import parse_version
 
 from skrub import (
+    SquashingScaler,
     StringEncoder,
     TableVectorizer,
     ToCategorical,
@@ -56,7 +57,7 @@ def test_linear_learner():
     assert isinstance(tv.high_cardinality, StringEncoder)
     assert isinstance(tv.low_cardinality, OneHotEncoder)
     assert isinstance(imputer, SimpleImputer)
-    assert isinstance(scaler, StandardScaler)
+    assert isinstance(scaler, SquashingScaler)
     assert tv.datetime.periodic_encoding == "spline"