FIX Treat gradient boosting categoricals outside the bounds as unknown during predict (#24283)

thomasjpfan · web-flow · commit 072b481600c4 · 2022-09-30T14:58:49.000+02:00
diff --git a/doc/whats_new/v1.2.rst b/doc/whats_new/v1.2.rst
@@ -262,6 +262,11 @@ Changelog
   NaN in feature importance when fitted with very small sample weight.
   :pr:`20415` by :user:`Zhehao Liu <MaxwellLZH>`.
 
+- |Fix| :class:`ensemble.HistGradientBoostingClassifier` and
+  :class:`ensemble.HistGradientBoostingRegressor` no longer error when predicting
+  on categories encoded as negative values and instead consider them a member
+  of the "missing category". :pr:`24283` by `Thomas Fan`_.
+
 - |API| Rename the constructor parameter `base_estimator` to `estimator` in
   the following classes:
   :class:`ensemble.BaggingClassifier`,
diff --git a/sklearn/ensemble/_hist_gradient_boosting/_predictor.pyx b/sklearn/ensemble/_hist_gradient_boosting/_predictor.pyx
@@ -60,7 +60,10 @@ cdef inline Y_DTYPE_C _predict_one_from_raw_data(
             else:
                 node_idx = node.right
         elif node.is_categorical:
-            if in_bitset_2d_memoryview(
+            if data_val < 0:
+                # data_val is not in the accepted range, so it is treated as missing value
+                node_idx = node.left if node.missing_go_to_left else node.right
+            elif in_bitset_2d_memoryview(
                     raw_left_cat_bitsets,
                     <X_BINNED_DTYPE_C>data_val,
                     node.bitset_idx):
diff --git a/sklearn/ensemble/_hist_gradient_boosting/gradient_boosting.py b/sklearn/ensemble/_hist_gradient_boosting/gradient_boosting.py
@@ -1176,6 +1176,8 @@ class HistGradientBoostingRegressor(RegressorMixin, BaseHistGradientBoosting):
 
         For each categorical feature, there must be at most `max_bins` unique
         categories, and each categorical value must be in [0, max_bins -1].
+        During prediction, categories encoded as a negative value are treated as
+        missing values.
 
         Read more in the :ref:`User Guide <categorical_support_gbdt>`.
 
@@ -1488,6 +1490,8 @@ class HistGradientBoostingClassifier(ClassifierMixin, BaseHistGradientBoosting):
 
         For each categorical feature, there must be at most `max_bins` unique
         categories, and each categorical value must be in [0, max_bins -1].
+        During prediction, categories encoded as a negative value are treated as
+        missing values.
 
         Read more in the :ref:`User Guide <categorical_support_gbdt>`.
 
diff --git a/sklearn/ensemble/_hist_gradient_boosting/tests/test_gradient_boosting.py b/sklearn/ensemble/_hist_gradient_boosting/tests/test_gradient_boosting.py
@@ -1178,3 +1178,28 @@ def test_class_weights():
         clf_balanced.decision_function(X_imb),
         clf_sample_weight.decision_function(X_imb),
     )
+
+
+def test_unknown_category_that_are_negative():
+    """Check that unknown categories that are negative does not error.
+
+    Non-regression test for #24274.
+    """
+    rng = np.random.RandomState(42)
+    n_samples = 1000
+    X = np.c_[rng.rand(n_samples), rng.randint(4, size=n_samples)]
+    y = np.zeros(shape=n_samples)
+    y[X[:, 1] % 2 == 0] = 1
+
+    hist = HistGradientBoostingRegressor(
+        random_state=0,
+        categorical_features=[False, True],
+        max_iter=10,
+    ).fit(X, y)
+
+    # Check that negative values from the second column are treated like a
+    # missing category
+    X_test_neg = np.asarray([[1, -2], [3, -4]])
+    X_test_nan = np.asarray([[1, np.nan], [3, np.nan]])
+
+    assert_allclose(hist.predict(X_test_neg), hist.predict(X_test_nan))
