prepare release 2.6

CHANGELOG.md

@@ -1,7 +1,23 @@
 unreleased
 ==========
-* added gray encoder
 
+v2.6.0
+======
+* added gray encoder
+* added thermometer / rank-hot encoder
+* introduce compatibility with sklearn 1.2
+  * compatibility with `feature_names_out_`
+  * remove boston housing dataset
+  * drop support for dataframes with non-homogenous data types in column names (i.e. having both string and integer column names)
+* improve performance of hashing encoder
+* improve catboost documentation
+* fix inverse transform in baseN with special character column names (issue 392)
+* fix inverse transform of ordinal encoder with custom mapping (issue 202)
+* fix re-fittable polynomial wrapper (issue 313)
+* fix numerical stability for target encoding (issue 377)
+* change default parameters of target encoding (issue 327)
+* drop support for sklearn 0.x
+
 v2.5.1.post0
 ============
 * fix pypi sdist

README.md

@@ -26,6 +26,7 @@ __Unsupervised:__
  * Helmert Contrast [2][3]
  * Ordinal [2][3]
  * One-Hot [2][3]
+ * Rank Hot [15]
  * Polynomial Contrast [2][3]
  * Sum Contrast [2][3]
 
@@ -149,3 +150,4 @@ References
  12. Andrew Gelman and Jennifer Hill (2006). Data Analysis Using Regression and Multilevel/Hierarchical Models. From https://faculty.psau.edu.sa/filedownload/doc-12-pdf-a1997d0d31f84d13c1cdc44ac39a8f2c-original.pdf
  13. Carlos Mougan, David Masip, Jordi Nin and Oriol Pujol (2021). Quantile Encoder: Tackling High Cardinality Categorical Features in Regression Problems. https://link.springer.com/chapter/10.1007%2F978-3-030-85529-1_14
  14. Gray Encoding. From https://en.wikipedia.org/wiki/Gray_code 
+ 15. Jacob Buckman, Aurko Roy, Colin Raffel, Ian Goodfellow: Thermometer Encoding: One Hot Way To Resist Adversarial Examples. From https://openreview.net/forum?id=S18Su--CW

category_encoders/__init__.py

@@ -27,7 +27,7 @@
 from category_encoders.glmm import GLMMEncoder
 from category_encoders.quantile_encoder import QuantileEncoder, SummaryEncoder
 
-__version__ = '2.5.1.post0'
+__version__ = '2.6.0'
 
 __author__ = "willmcginnis", "cmougan", "paulwestenthanner"
 

category_encoders/target_encoder.py

@@ -4,7 +4,6 @@
 from scipy.special import expit
 from category_encoders.ordinal import OrdinalEncoder
 import category_encoders.utils as util
-import warnings
 
 __author__ = 'chappers'
 
@@ -44,10 +43,10 @@ class TargetEncoder(util.BaseEncoder, util.SupervisedTransformerMixin):
         The value must be strictly bigger than 0. Higher values mean a flatter S-curve (see min_samples_leaf).
     hierarchy: dict or dataframe
         A dictionary or a dataframe to define the hierarchy for mapping.
-        
+
         If a dictionary, this contains a dict of columns to map into hierarchies.  Dictionary key(s) should be the column name from X
         which requires mapping.  For multiple hierarchical maps, this should be a dictionary of dictionaries.
-        
+
         If dataframe: a dataframe defining columns to be used for the hierarchies.  Column names must take the form:
             HIER_colA_1, ... HIER_colA_N, HIER_colB_1, ... HIER_colB_M, ...
         where [colA, colB, ...] are given columns in cols list.  
@@ -111,20 +110,12 @@ class TargetEncoder(util.BaseEncoder, util.SupervisedTransformerMixin):
     encoding_relation = util.EncodingRelation.ONE_TO_ONE
 
     def __init__(self, verbose=0, cols=None, drop_invariant=False, return_df=True, handle_missing='value',
-                 handle_unknown='value', min_samples_leaf=1, smoothing=1.0, hierarchy=None):
+                 handle_unknown='value', min_samples_leaf=20, smoothing=10, hierarchy=None):
         super().__init__(verbose=verbose, cols=cols, drop_invariant=drop_invariant, return_df=return_df,
                          handle_unknown=handle_unknown, handle_missing=handle_missing)
         self.ordinal_encoder = None
         self.min_samples_leaf = min_samples_leaf
-        if min_samples_leaf == 1:
-            warnings.warn("Default parameter min_samples_leaf will change in version 2.6."
-                          "See https://github.com/scikit-learn-contrib/category_encoders/issues/327",
-                          category=FutureWarning)
         self.smoothing = smoothing
-        if smoothing == 1.0:
-            warnings.warn("Default parameter smoothing will change in version 2.6."
-                          "See https://github.com/scikit-learn-contrib/category_encoders/issues/327",
-                          category=FutureWarning)
         self.mapping = None
         self._mean = None
         if isinstance(hierarchy, (dict, pd.DataFrame)) and cols is None:
@@ -203,7 +194,7 @@ def fit_target_encoding(self, X, y):
             col = switch.get('col')
             if 'HIER_' not in str(col):
                 values = switch.get('mapping')
-                
+
                 scalar = prior
                 if (isinstance(self.hierarchy, dict) and col in self.hierarchy) or \
                                     (isinstance(self.hierarchy, pd.DataFrame)):
@@ -222,7 +213,6 @@ def fit_target_encoding(self, X, y):
                 smoove = self._weighting(stats['count'])
 
                 smoothing = scalar * (1 - smoove) + stats['mean'] * smoove
-                smoothing[stats['count'] == 1] = scalar
 
                 if self.handle_unknown == 'return_nan':
                     smoothing.loc[-1] = np.nan

tests/test_encoders.py

@@ -381,7 +381,7 @@ def test_preserve_names(self):
     def test_unique_column_is_not_predictive(self):
         # @ToDo not sure how useful this test is. TargetEncoders set the value to the default if there is only
         #  one category but they probably should not. See discussion in issue 327
-        test_encoders = ['LeaveOneOutEncoder', 'TargetEncoder', 'WOEEncoder', 'MEstimateEncoder',
+        test_encoders = ['LeaveOneOutEncoder', 'WOEEncoder', 'MEstimateEncoder',
                          'JamesSteinEncoder', 'CatBoostEncoder', 'GLMMEncoder']
         for encoder_name in test_encoders:
             enc = getattr(encoders, encoder_name)()

tests/test_target_encoder.py

@@ -109,11 +109,6 @@ def test_target_encoder_fit_transform_HaveNanValue_ExpectCorrectValueInResult(se
         self.assertAlmostEqual(0.4125, values[2], delta=1e-4)
         self.assertEqual(0.5, values[3])
 
-    def test_target_encoder_noncontiguous_index(self):
-        data = pd.DataFrame({'x': ['a', 'b', np.nan, 'd', 'e'], 'y': range(5)}).dropna()
-        result = encoders.TargetEncoder(cols=['x']).fit_transform(data[['x']], data['y'])
-        self.assertTrue(np.allclose(result, 2.0))
-
     def test_HandleMissingIsValueAndNanInTest_ExpectMean(self):
         df = pd.DataFrame({
             'color': ["a", "a", "a", "b", "b", "b"],
@@ -175,7 +170,7 @@ def test_hierarchical_smoothing_multi(self):
         self.assertAlmostEqual(0.3248, values[5], delta=1e-4)
         self.assertAlmostEqual(0.6190, values[11], delta=1e-4)
         self.assertAlmostEqual(0.1309, values[13], delta=1e-4)
-        self.assertAlmostEqual(0.7381, values[15], delta=1e-4)
+        self.assertAlmostEqual(0.8370, values[15], delta=1e-4)
 
     def test_hierarchical_part_named_cols(self):
 
@@ -299,10 +294,10 @@ def test_hierarchy_multi_level(self):
         values = result['Animal'].values
         self.assertAlmostEqual(0.6261, values[0], delta=1e-4)
         self.assertAlmostEqual(0.9065, values[2], delta=1e-4)
-        self.assertAlmostEqual(0.4107, values[5], delta=1e-4)
+        self.assertAlmostEqual(0.2556, values[5], delta=1e-4)
         self.assertAlmostEqual(0.3680, values[8], delta=1e-4)
         self.assertAlmostEqual(0.4626, values[11], delta=1e-4)
-        self.assertAlmostEqual(0.2466, values[13], delta=1e-4)
+        self.assertAlmostEqual(0.1535, values[13], delta=1e-4)
         self.assertAlmostEqual(0.4741, values[14], delta=1e-4)
 
     def test_hierarchy_columnwise_compass(self):
@@ -330,7 +325,7 @@ def test_hierarchy_columnwise_postcodes(self):
         result = enc.fit_transform(X[cols], y)
 
         values = result['postcode'].values
-        self.assertAlmostEqual(0.7506, values[0], delta=1e-4)
+        self.assertAlmostEqual(0.8448, values[0], delta=1e-4)
 
 
     def test_hierarchy_columnwise_missing_level(self):