ENH Support unknown_value=np.nan in OrdinalEncoder (scikit-learn#18406)

Author: NicolasHug

doc/whats_new/v0.24.rst

@@ -542,7 +542,8 @@ Changelog
   ``use_encoded_value`` option, along with a new ``unknown_value`` parameter,
   to :class:`preprocessing.OrdinalEncoder` to allow unknown categories during
   transform and set the encoded value of the unknown categories.
-  :pr:`17406` by :user:`Felix Wick <FelixWick>`.
+  :pr:`17406` by :user:`Felix Wick <FelixWick>` and :pr:`18406` by
+  `Nicolas Hug`_.
 
 - |Feature| Add ``clip`` parameter to :class:`preprocessing.MinMaxScaler`,
   which clips the transformed values of test data to ``feature_range``.

sklearn/preprocessing/_encoders.py

@@ -7,7 +7,7 @@
 import numbers
 
 from ..base import BaseEstimator, TransformerMixin
-from ..utils import check_array
+from ..utils import check_array, is_scalar_nan
 from ..utils.validation import check_is_fitted
 from ..utils.validation import _deprecate_positional_args
 
@@ -631,11 +631,12 @@ class OrdinalEncoder(_BaseEncoder):
 
         .. versionadded:: 0.24
 
-    unknown_value : int, default=None
+    unknown_value : int or np.nan, default=None
         When the parameter handle_unknown is set to 'use_encoded_value', this
         parameter is required and will set the encoded value of unknown
         categories. It has to be distinct from the values used to encode any of
-        the categories in `fit`.
+        the categories in `fit`. If set to np.nan, the `dtype` parameter must
+        be a float dtype.
 
         .. versionadded:: 0.24
 
@@ -699,13 +700,21 @@ def fit(self, X, y=None):
         self
         """
         if self.handle_unknown == 'use_encoded_value':
-            if not isinstance(self.unknown_value, numbers.Integral):
-                raise TypeError(f"unknown_value should be an integer when "
-                                f"`handle_unknown is 'use_encoded_value'`, "
+            if is_scalar_nan(self.unknown_value):
+                if np.dtype(self.dtype).kind != 'f':
+                    raise ValueError(
+                        f"When unknown_value is np.nan, the dtype "
+                        "parameter should be "
+                        f"a float dtype. Got {self.dtype}."
+                    )
+            elif not isinstance(self.unknown_value, numbers.Integral):
+                raise TypeError(f"unknown_value should be an integer or "
+                                f"np.nan when "
+                                f"handle_unknown is 'use_encoded_value', "
                                 f"got {self.unknown_value}.")
         elif self.unknown_value is not None:
             raise TypeError(f"unknown_value should only be set when "
-                            f"`handle_unknown is 'use_encoded_value'`, "
+                            f"handle_unknown is 'use_encoded_value', "
                             f"got {self.unknown_value}.")
 
         self._fit(X)
@@ -735,11 +744,12 @@ def transform(self, X):
             Transformed input.
         """
         X_int, X_mask = self._transform(X, handle_unknown=self.handle_unknown)
+        X_trans = X_int.astype(self.dtype, copy=False)
 
         # create separate category for unknown values
         if self.handle_unknown == 'use_encoded_value':
-            X_int[~X_mask] = self.unknown_value
-        return X_int.astype(self.dtype, copy=False)
+            X_trans[~X_mask] = self.unknown_value
+        return X_trans
 
     def inverse_transform(self, X):
         """

sklearn/preprocessing/tests/test_encoders.py

@@ -589,21 +589,21 @@ def test_ordinal_encoder_handle_unknowns_raise():
     X = np.array([['a', 'x'], ['b', 'y']], dtype=object)
 
     enc = OrdinalEncoder(handle_unknown='use_encoded_value')
-    msg = ("unknown_value should be an integer when `handle_unknown is "
-           "'use_encoded_value'`, got None.")
+    msg = ("unknown_value should be an integer or np.nan when handle_unknown "
+           "is 'use_encoded_value', got None.")
     with pytest.raises(TypeError, match=msg):
         enc.fit(X)
 
     enc = OrdinalEncoder(unknown_value=-2)
-    msg = ("unknown_value should only be set when `handle_unknown is "
-           "'use_encoded_value'`, got -2.")
+    msg = ("unknown_value should only be set when handle_unknown is "
+           "'use_encoded_value', got -2.")
     with pytest.raises(TypeError, match=msg):
         enc.fit(X)
 
     enc = OrdinalEncoder(handle_unknown='use_encoded_value',
                          unknown_value='bla')
-    msg = ("unknown_value should be an integer when `handle_unknown is "
-           "'use_encoded_value'`, got bla.")
+    msg = ("unknown_value should be an integer or np.nan when handle_unknown "
+           "is 'use_encoded_value', got bla.")
     with pytest.raises(TypeError, match=msg):
         enc.fit(X)
 
@@ -614,6 +614,30 @@ def test_ordinal_encoder_handle_unknowns_raise():
         enc.fit(X)
 
 
+def test_ordinal_encoder_handle_unknowns_nan():
+    # Make sure unknown_value=np.nan properly works
+
+    enc = OrdinalEncoder(handle_unknown='use_encoded_value',
+                         unknown_value=np.nan)
+
+    X_fit = np.array([[1], [2], [3]])
+    enc.fit(X_fit)
+    X_trans = enc.transform([[1], [2], [4]])
+    assert_array_equal(X_trans, [[0], [1], [np.nan]])
+
+
+def test_ordinal_encoder_handle_unknowns_nan_non_float_dtype():
+    # Make sure an error is raised when unknown_value=np.nan and the dtype
+    # isn't a float dtype
+    enc = OrdinalEncoder(handle_unknown='use_encoded_value',
+                         unknown_value=np.nan, dtype=int)
+
+    X_fit = np.array([[1], [2], [3]])
+    with pytest.raises(ValueError,
+                       match="dtype parameter should be a float dtype"):
+        enc.fit(X_fit)
+
+
 def test_ordinal_encoder_raise_categories_shape():
 
     X = np.array([['Low', 'Medium', 'High', 'Medium', 'Low']], dtype=object).T