Disable TargetEncoding for Classification

There are a few bugs that prevent the TargetEncoder from working as
intended (see
scikit-learn-contrib/category_encoders#182):
 - It does not work with a `categorical` series as target
 - It does not work for multi-class classification

gama/gama.py

@@ -479,8 +479,13 @@ def fit(
         ):
             x, self._y = format_x_y(x, y)
             self._inferred_dtypes = x.dtypes
-            self._x, self._basic_encoding_pipeline = basic_encoding(x)
-            self._fixed_pipeline_extension = basic_pipeline_extension(self._x)
+            is_classification = hasattr(self, "_label_encoder")
+            self._x, self._basic_encoding_pipeline = basic_encoding(
+                x, is_classification
+            )
+            self._fixed_pipeline_extension = basic_pipeline_extension(
+                self._x, is_classification
+            )
             self._operator_set._safe_compile = partial(
                 compile_individual, preprocessing_steps=self._fixed_pipeline_extension
             )

gama/utilities/preprocessing.py

@@ -42,34 +42,44 @@ def select_categorical_columns(
                 yield column
 
 
-def basic_encoding(x: pd.DataFrame):
+def basic_encoding(x: pd.DataFrame, is_classification: bool):
     """ Perform 'basic' encoding of categorical features.
 
      Specifically, perform:
       - Ordinal encoding for features with 2 or fewer unique values.
       - One hot encoding for features with at most 10 unique values.
+      - Ordinal encoding for features with 11+ unique values, if y is categorical.
      """
-    binary_features = list(select_categorical_columns(x, max_f=2))
+    ord_features = list(select_categorical_columns(x, max_f=2))
+    if is_classification:
+        ord_features.extend(select_categorical_columns(x, min_f=11))
     leq_10_features = list(select_categorical_columns(x, min_f=3, max_f=10))
 
-    encoding_pipeline = Pipeline(
-        steps=[
-            ("ord-enc", ce.OrdinalEncoder(cols=binary_features, drop_invariant=True)),
-            ("oh-enc", ce.OneHotEncoder(cols=leq_10_features, handle_missing="ignore")),
-        ]
-    )
-    x_enc = encoding_pipeline.fit_transform(x, y=None)  # Is this allowed?
+    encoding_steps = [
+        ("ord-enc", ce.OrdinalEncoder(cols=ord_features, drop_invariant=True)),
+        ("oh-enc", ce.OneHotEncoder(cols=leq_10_features, handle_missing="ignore")),
+    ]
+    encoding_pipeline = Pipeline(encoding_steps)
+    x_enc = encoding_pipeline.fit_transform(x, y=None)  # Is this dangerous?
     return x_enc, encoding_pipeline
 
 
-def basic_pipeline_extension(x: pd.DataFrame) -> List[Tuple[str, TransformerMixin]]:
+def basic_pipeline_extension(
+    x: pd.DataFrame, is_classification: bool
+) -> List[Tuple[str, TransformerMixin]]:
     """ Define a TargetEncoder and SimpleImputer.
 
-    TargetEncoding is will encode categorical features with more than 10 unique values.
-    SimpleImputer imputes with the median.
+    TargetEncoding is will encode categorical features with more than 10 unique values,
+    if y is not categorical. SimpleImputer imputes with the median.
     """
-    many_factor_features = list(select_categorical_columns(x, min_f=11))
-    return [
-        ("target_enc", ce.TargetEncoder(cols=many_factor_features)),
-        ("imputation", SimpleImputer(strategy="median")),
-    ]
+    # These steps need to be in the pipeline because they need to be trained each fold.
+    extension_steps = []
+    if not is_classification:
+        # TargetEncoder is broken with categorical target
+        many_factor_features = list(select_categorical_columns(x, min_f=11))
+        extension_steps.append(
+            ("target_enc", ce.TargetEncoder(cols=many_factor_features))
+        )
+    extension_steps.append(("imputation", SimpleImputer(strategy="median")))
+
+    return extension_steps