Expose classes_/predict_proba on calibrated classifiers for global explainers

DashAI/back/models/scikit_learn/linear_svc_classifier.py

@@ -295,6 +295,40 @@ def __init__(self, **kwargs) -> None:
     def __sklearn_is_fitted__(self) -> bool:
         return self._calibrated is not None
 
+    @property
+    def classes_(self):
+        """Expose the calibrated model's classes for sklearn compatibility.
+
+        sklearn utilities (e.g. permutation_importance, partial_dependence)
+        read ``estimator.classes_`` directly on any object tagged as a
+        classifier, regardless of which prediction method they end up
+        calling. Since this wrapper fits ``self._calibrated`` instead of
+        ``self``, that attribute must be proxied explicitly.
+        """
+        from sklearn.exceptions import NotFittedError
+
+        if self._calibrated is None:
+            raise NotFittedError(
+                f"This {self.__class__.__name__} instance is not fitted yet. "
+                "Call 'train' with appropriate arguments before using this estimator."
+            )
+        return self._calibrated.classes_
+
+    def predict_proba(self, x_pred) -> "ndarray":  # noqa: F821
+        """Return class-probability matrix using the calibrated model.
+
+        Parameters
+        ----------
+        x_pred : DashAIDataset or pd.DataFrame
+            Input data.
+
+        Returns
+        -------
+        np.ndarray
+            Class probability matrix.
+        """
+        return self.predict(x_pred)
+
     def train(self, x_train, y_train, x_validation=None, y_validation=None):
         """Train using CalibratedClassifierCV to expose predict_proba.
 

DashAI/back/models/scikit_learn/sgd_classifier.py

@@ -325,6 +325,40 @@ def __init__(self, **kwargs) -> None:
     def __sklearn_is_fitted__(self) -> bool:
         return self._calibrated is not None
 
+    @property
+    def classes_(self):
+        """Expose the calibrated model's classes for sklearn compatibility.
+
+        sklearn utilities (e.g. permutation_importance, partial_dependence)
+        read ``estimator.classes_`` directly on any object tagged as a
+        classifier, regardless of which prediction method they end up
+        calling. Since this wrapper fits ``self._calibrated`` instead of
+        ``self``, that attribute must be proxied explicitly.
+        """
+        from sklearn.exceptions import NotFittedError
+
+        if self._calibrated is None:
+            raise NotFittedError(
+                f"This {self.__class__.__name__} instance is not fitted yet. "
+                "Call 'train' with appropriate arguments before using this estimator."
+            )
+        return self._calibrated.classes_
+
+    def predict_proba(self, x_pred) -> "ndarray":  # noqa: F821
+        """Return class-probability matrix using the calibrated model.
+
+        Parameters
+        ----------
+        x_pred : DashAIDataset or pd.DataFrame
+            Input data.
+
+        Returns
+        -------
+        np.ndarray
+            Class probability matrix.
+        """
+        return self.predict(x_pred)
+
     def train(self, x_train, y_train, x_validation=None, y_validation=None):
         """Train using CalibratedClassifierCV to guarantee predict_proba availability.
 

tests/back/explainers/test_explainers.py

@@ -21,6 +21,8 @@
 from DashAI.back.models.scikit_learn.decision_tree_classifier import (
     DecisionTreeClassifier,
 )
+from DashAI.back.models.scikit_learn.linear_svc_classifier import LinearSVCClassifier
+from DashAI.back.models.scikit_learn.sgd_classifier import SGDClassifier
 from DashAI.back.types.categorical import Categorical
 from DashAI.back.types.utils import save_types_in_arrow_metadata
 from DashAI.back.types.value_types import Float
@@ -114,6 +116,56 @@ def trained_model(dataset):
     return model
 
 
+@pytest.fixture(scope="module", params=[LinearSVCClassifier, SGDClassifier])
+def trained_calibrated_model(request, dataset):
+    """Models that internally calibrate via CalibratedClassifierCV.
+
+    Regression coverage for the bug where these wrappers didn't proxy
+    ``classes_``/``predict_proba`` onto themselves, breaking any explainer
+    that relies on sklearn's classifier introspection (e.g. PFI, PDP).
+    """
+    x, y = dataset
+    model = request.param()
+    model.train(x["train"], y["train"])
+    return model
+
+
+def test_permutation_feature_importance_calibrated_model(
+    trained_calibrated_model: BaseModel, dataset: DatasetDict
+):
+    explainer = PermutationFeatureImportance(
+        trained_calibrated_model,
+        scoring="accuracy",
+        n_repeats=5,
+        random_state=None,
+        max_samples_fraction=1.0,
+    )
+    explanation = explainer.explain(copy.deepcopy(dataset))
+
+    assert all(
+        key in explanation
+        for key in ["features", "importances_mean", "importances_std"]
+    )
+    for values in explanation.values():
+        assert len(values) == len(INPUT_COLUMNS)
+
+
+def test_partial_dependence_calibrated_model(
+    trained_calibrated_model: BaseModel, dataset
+):
+    explainer = PartialDependence(
+        trained_calibrated_model,
+        grid_resolution=50,
+        lower_percentile=0.01,
+        upper_percentile=0.99,
+    )
+    explanation = explainer.explain(copy.deepcopy(dataset))
+
+    metadata = explanation.pop("metadata")
+    assert set(metadata["target_names"]) == set(TARGETS)
+    assert len(explanation) == len(INPUT_COLUMNS)
+
+
 def test_partial_dependence(trained_model: BaseModel, dataset):
     parameters = {
         "grid_resolution": 50,