scikit-learn · jeremiedbb · Dec 4, 2024 · Nov 29, 2024 · Nov 29, 2024 · Nov 29, 2024
diff --git a/sklearn/neighbors/_base.py b/sklearn/neighbors/_base.py
@@ -709,6 +709,8 @@ def __sklearn_tags__(self):
         tags = super().__sklearn_tags__()
         # For cross-validation routines to split data correctly
         tags.input_tags.pairwise = self.metric == "precomputed"
+        # when input is precomputed metric values, all those values need to be positive
+        tags.input_tags.positive_only = tags.input_tags.pairwise
         tags.input_tags.allow_nan = self.metric == "nan_euclidean"
         return tags
 

diff --git a/sklearn/utils/_tags.py b/sklearn/utils/_tags.py
@@ -58,6 +58,10 @@ class InputTags:
         Specifically, this tag is used by
         `sklearn.utils.metaestimators._safe_split` to slice rows and
         columns.
+
+        Note that if setting this tag to ``True`` means the estimator can take only
+        positive values, the `positive_only` tag must reflect it and also be set to
+        ``True``.
     """
 
     one_d_array: bool = False

diff --git a/sklearn/utils/estimator_checks.py b/sklearn/utils/estimator_checks.py
@@ -148,6 +148,7 @@ def _yield_api_checks(estimator):
     yield check_do_not_raise_errors_in_init_or_set_params
     yield check_n_features_in_after_fitting
     yield check_mixin_order
+    yield check_positive_only_tag_during_fit
 
 
 def _yield_checks(estimator):
@@ -3899,6 +3900,39 @@ def _enforce_estimator_tags_X(estimator, X, X_test=None, kernel=linear_kernel):
     return X_res
 
 
+@ignore_warnings(category=FutureWarning)
+def check_positive_only_tag_during_fit(name, estimator_orig):
+    """Test that the estimator correctly sets the tags.input_tags.positive_only
+
+    If the tag is False, the estimator should accept negative input regardless of the
+    tags.input_tags.pairwise flag.
+    """
+    estimator = clone(estimator_orig)
+    tags = get_tags(estimator)
+
+    X, y = load_iris(return_X_y=True)
+    y = _enforce_estimator_tags_y(estimator, y)
+    set_random_state(estimator, 0)
+    X = _enforce_estimator_tags_X(estimator, X)
+    X -= X.mean()
+
+    if tags.input_tags.positive_only:
+        with raises(ValueError, match="Negative values in data"):
+            estimator.fit(X, y)
+    else:
+        # This should pass
+        try:
+            estimator.fit(X, y)
+        except Exception as e:
+            err_msg = (
+                f"Estimator {repr(name)} raised {e.__class__.__name__} unexpectedly."
+                " This happens when passing negative input values as X."
+                " If negative values are not supported for this estimator instance,"
+                " then the tags.input_tags.positive_only tag needs to be set to True."
+            )
+            raise AssertionError(err_msg) from e
+
+
 @ignore_warnings(category=FutureWarning)
 def check_non_transformer_estimators_n_iter(name, estimator_orig):
     # Test that estimators that are not transformers with a parameter

diff --git a/sklearn/utils/tests/test_estimator_checks.py b/sklearn/utils/tests/test_estimator_checks.py
@@ -85,6 +85,7 @@
     check_outlier_contamination,
     check_outlier_corruption,
     check_parameters_default_constructible,
+    check_positive_only_tag_during_fit,
     check_regressor_data_not_an_array,
     check_requires_y_none,
     check_sample_weights_pandas_series,
@@ -509,7 +510,7 @@ class RequiresPositiveXRegressor(LinearRegression):
     def fit(self, X, y):
         X, y = validate_data(self, X, y, multi_output=True)
         if (X < 0).any():
-            raise ValueError("negative X values not supported!")
+            raise ValueError("Negative values in data passed to X.")
         return super().fit(X, y)
 
     def __sklearn_tags__(self):
@@ -1600,3 +1601,18 @@ def fit(self, X, y=None):
     msg = "TransformerMixin comes before/left side of BaseEstimator"
     with raises(AssertionError, match=re.escape(msg)):
         check_mixin_order("BadEstimator", BadEstimator())
+
+
+def test_check_positive_only_tag_during_fit():
+    class RequiresPositiveXBadTag(RequiresPositiveXRegressor):
+        def __sklearn_tags__(self):
+            tags = super().__sklearn_tags__()
+            tags.input_tags.positive_only = False
+            return tags
+
+    with raises(
+        AssertionError, match="This happens when passing negative input values as X."
+    ):
+        check_positive_only_tag_during_fit(
+            "RequiresPositiveXBadTag", RequiresPositiveXBadTag()
+        )