Fix multiclass LinearSVC fallback paths and close #256

Author: oegedijk

RELEASE_NOTES.md

@@ -10,6 +10,8 @@
 - Fix issue #294: align multiclass `model_output='logodds'` semantics across Prediction Box and Contributions Plot by using per-class raw margins for multiclass logodds displays.
 - Fix multiclass PDP highlight predictions in logodds mode to use the same raw-margin scale as SHAP contributions.
 - Fix XGBoost multiclass decision-path summary wording to display `prediction (logodds)` when explainer `model_output='logodds'`.
+- Fix issue #256: add robust multiclass probability fallback for classifiers that expose `decision_function` but not `predict_proba` (e.g. `LinearSVC`), and use it consistently across kernel SHAP, prediction helpers, PDP, and permutation scorer paths.
+- Prevent multiclass class-count mismatches when user-provided/broken `predict_proba` outputs do not match model class count by falling back to `decision_function`-based probabilities.
 
 ### Tests
 - Add regression tests for LightGBM with string categorical features covering dashboard initialization, `get_shap_row(...)`, unseen categorical values in `X_row`, and regression dashboard initialization.
@@ -18,6 +20,8 @@
 - Add regression tests for issue #294 covering multiclass logodds consistency across prediction table, contributions, PDP highlight predictions, and XGBoost decision-path summaries.
 - Add pipeline tests for transformed feature-name cleanup (`strip_pipeline_prefix`, `feature_name_fn`) and pipeline categorical grouping autodetection.
 - Add explainer-method unit tests for binary-like onehot detection, transformed feature-name deduping, inferred pipeline cats, and pipeline extraction warning text.
+- Add regression tests for issue #256 covering multiclass `LinearSVC` with kernel SHAP, PDP, and permutation-importances flows using `decision_function` fallback.
+- Add guard tests to confirm multiclass `predict_proba` models (logistic regression) keep working for PDP and permutation-importances paths.
 
 ### Improvements
 - Add pipeline feature-name cleanup options: `strip_pipeline_prefix=True` and `feature_name_fn=...` for sklearn/imblearn pipeline transformed output columns.

explainerdashboard/explainer_methods.py

@@ -200,6 +200,87 @@ def _convert_elem(elem):
     return pred_array
 
 
+def _decision_scores_to_probas(decision_scores, n_labels=None):
+    """Map decision_function outputs to probability-like class scores."""
+    scores = np.asarray(decision_scores)
+    if scores.ndim == 0:
+        scores = scores.reshape(1)
+    if scores.ndim == 1 and n_labels and n_labels > 2 and scores.shape[0] == n_labels:
+        scores = scores.reshape(1, -1)
+
+    if scores.ndim == 1:
+        clipped = np.clip(scores.astype("float64"), -709, 709)
+        pos_probs = 1.0 / (1.0 + np.exp(-clipped))
+        return np.column_stack([1.0 - pos_probs, pos_probs])
+
+    if scores.ndim == 2:
+        if scores.shape[1] == 1:
+            clipped = np.clip(scores[:, 0].astype("float64"), -709, 709)
+            pos_probs = 1.0 / (1.0 + np.exp(-clipped))
+            return np.column_stack([1.0 - pos_probs, pos_probs])
+
+        shifted = scores - np.max(scores, axis=1, keepdims=True)
+        exp_scores = np.exp(shifted)
+        denom = np.sum(exp_scores, axis=1, keepdims=True)
+        return exp_scores / np.clip(denom, np.finfo("float64").tiny, None)
+
+    raise ValueError(
+        f"Unexpected decision_function output shape {scores.shape}. "
+        "Expected 1D or 2D scores."
+    )
+
+
+def _predict_proba_with_fallback(model, model_input, n_labels=None):
+    """Return per-class probabilities, with decision_function fallback."""
+    pred_probas = None
+    predict_error = None
+
+    if hasattr(model, "predict_proba"):
+        try:
+            pred_raw = model.predict_proba(model_input)
+            pred_raw = _ensure_numeric_predictions(pred_raw)
+            pred_probas = np.asarray(pred_raw, dtype="float64")
+        except Exception as e:
+            predict_error = e
+
+    if pred_probas is not None:
+        if pred_probas.ndim == 1:
+            if n_labels == 2:
+                pred_probas = np.column_stack([1.0 - pred_probas, pred_probas])
+            else:
+                pred_probas = None
+        elif pred_probas.ndim != 2:
+            pred_probas = None
+
+        if (
+            pred_probas is not None
+            and n_labels is not None
+            and pred_probas.shape[1] != n_labels
+        ):
+            pred_probas = None
+
+    if pred_probas is None and hasattr(model, "decision_function"):
+        decision_scores_raw = model.decision_function(model_input)
+        decision_scores_raw = _ensure_numeric_predictions(decision_scores_raw)
+        pred_probas = _decision_scores_to_probas(decision_scores_raw, n_labels=n_labels)
+
+    if pred_probas is None:
+        if predict_error is not None:
+            raise ValueError(
+                "Could not compute class probabilities from model.predict_proba(...)."
+            ) from predict_error
+        raise ValueError(
+            "Could not compute class probabilities: model has neither a working "
+            "predict_proba(...) nor decision_function(...)."
+        )
+
+    if n_labels is not None and pred_probas.shape[1] != n_labels:
+        raise ValueError(
+            f"Expected {n_labels} class probabilities, got shape {pred_probas.shape}."
+        )
+    return pred_probas
+
+
 def get_multiclass_logodds_scores(model, model_input, n_labels):
     """Return per-class raw scores used as multiclass logodds/margins.
 
@@ -288,7 +369,16 @@ def _wrapped_scorer(estimator, X, y_true):
             if "__sklearn_tags__" in str(e):
                 # Model doesn't have __sklearn_tags__, call predict/predict_proba directly
                 if response_method == "predict_proba":
-                    y_pred = estimator.predict_proba(X)
+                    n_labels = (
+                        len(estimator.classes_)
+                        if hasattr(estimator, "classes_")
+                        else None
+                    )
+                    y_pred = _predict_proba_with_fallback(
+                        estimator,
+                        X,
+                        n_labels=n_labels,
+                    )
                 else:
                     y_pred = estimator.predict(X)
                     y_pred = _ensure_numeric_predictions(y_pred)
@@ -307,7 +397,14 @@ def _wrapped_scorer(estimator, X, y_true):
         # If scorer creation failed, use direct prediction
         if scorer is None:
             if response_method == "predict_proba":
-                y_pred = estimator.predict_proba(X)
+                n_labels = (
+                    len(estimator.classes_) if hasattr(estimator, "classes_") else None
+                )
+                y_pred = _predict_proba_with_fallback(
+                    estimator,
+                    X,
+                    n_labels=n_labels,
+                )
             else:
                 y_pred = estimator.predict(X)
                 y_pred = _ensure_numeric_predictions(y_pred)
@@ -1066,7 +1163,8 @@ def one_vs_all_metric(metric, pos_label, y_true, y_pred):
 
     def _scorer(clf, X, y):
         warnings.filterwarnings("ignore", category=UserWarning)
-        y_pred = clf.predict_proba(X)
+        n_labels = len(clf.classes_) if hasattr(clf, "classes_") else None
+        y_pred = _predict_proba_with_fallback(clf, X, n_labels=n_labels)
         warnings.filterwarnings("default", category=UserWarning)
         y_pred = _ensure_numeric_predictions(y_pred)
         y_pred = np.asarray(y_pred)
@@ -1440,7 +1538,10 @@ def _model_input(data):
     if is_classifier:
         first_row = _model_input(X_sample.iloc[[0]])
         warnings.filterwarnings("ignore", category=UserWarning)
-        n_labels = model.predict_proba(first_row).shape[1]
+        class_count = len(model.classes_) if hasattr(model, "classes_") else None
+        n_labels = _predict_proba_with_fallback(
+            model, first_row, n_labels=class_count
+        ).shape[1]
         warnings.filterwarnings("default", category=UserWarning)
         if multiclass:
             pdp_dfs = [pd.DataFrame() for i in range(n_labels)]
@@ -1474,9 +1575,11 @@ def _coerce_value(value, dtype):
         )
         if is_classifier:
             dtemp_model = _model_input(dtemp)
-            pred_probas_raw = model.predict_proba(dtemp_model)
-            pred_probas_raw = _ensure_numeric_predictions(pred_probas_raw)
-            pred_probas = np.asarray(pred_probas_raw).squeeze()
+            pred_probas = _predict_proba_with_fallback(
+                model,
+                dtemp_model,
+                n_labels=n_labels,
+            ).squeeze()
             if multiclass:
                 for i in range(n_labels):
                     pdp_dfs[i][grid_value] = pred_probas[:, i]

explainerdashboard/explainers.py

@@ -1058,9 +1058,11 @@ def get_col_value_plus_prediction(
             if self.is_classifier:
                 if pos_label is None:
                     pos_label = self.pos_label
-                pred_probas_raw = self.model.predict_proba(model_input)
-                pred_probas_raw = _ensure_numeric_predictions(pred_probas_raw)
-                pred_probas = np.asarray(pred_probas_raw).squeeze()
+                pred_probas = self._predict_proba_from_model(
+                    self.model,
+                    model_input,
+                    n_labels=len(self.labels),
+                ).squeeze()
                 if pred_probas.ndim > 1:
                     pred_probas = pred_probas[0]
                 prediction = pred_probas[pos_label].squeeze()
@@ -2843,17 +2845,10 @@ def __init__(
             auto_detect_pipeline_cats,
         )
 
-        assert hasattr(model, "predict_proba"), (
+        assert hasattr(model, "predict_proba") or hasattr(model, "decision_function"), (
             "for ClassifierExplainer, model should be a scikit-learn "
-            "compatible *classifier* model that has a predict_proba(...) "
-            f"method, so not a {type(model)}! If you are using e.g an SVM "
-            "with hinge loss (which does not support predict_proba), you "
-            "can try the following monkey patch:\n\n"
-            "import types\n"
-            "def predict_proba(self, X):\n"
-            "    pred = self.predict(X)\n"
-            "    return np.array([1-pred, pred]).T\n"
-            "model.predict_proba = types.MethodType(predict_proba, model)\n"
+            "compatible *classifier* model that has either predict_proba(...) "
+            f"or decision_function(...), so not a {type(model)}!"
         )
 
         self._params_dict = {
@@ -2965,23 +2960,112 @@ def y_binary(self, pos_label):
                 self._y_binaries = [self.y.values for i in range(len(self.labels))]
         return self._y_binaries[pos_label]
 
+    def _decision_scores_to_probas(self, decision_scores, n_labels=None):
+        """Map decision_function outputs to probability-like class scores."""
+        scores = np.asarray(decision_scores)
+        if scores.ndim == 0:
+            scores = scores.reshape(1)
+        if (
+            scores.ndim == 1
+            and n_labels
+            and n_labels > 2
+            and scores.shape[0] == n_labels
+        ):
+            scores = scores.reshape(1, -1)
+
+        if scores.ndim == 1:
+            clipped = np.clip(scores.astype("float64"), -709, 709)
+            pos_probs = 1.0 / (1.0 + np.exp(-clipped))
+            return np.column_stack([1.0 - pos_probs, pos_probs])
+
+        if scores.ndim == 2:
+            if scores.shape[1] == 1:
+                clipped = np.clip(scores[:, 0].astype("float64"), -709, 709)
+                pos_probs = 1.0 / (1.0 + np.exp(-clipped))
+                return np.column_stack([1.0 - pos_probs, pos_probs])
+
+            shifted = scores - np.max(scores, axis=1, keepdims=True)
+            exp_scores = np.exp(shifted)
+            denom = np.sum(exp_scores, axis=1, keepdims=True)
+            return exp_scores / np.clip(denom, np.finfo("float64").tiny, None)
+
+        raise ValueError(
+            f"Unexpected decision_function output shape {scores.shape}. "
+            "Expected 1D or 2D scores."
+        )
+
+    def _predict_proba_from_model(self, model, model_input, n_labels=None):
+        """Return per-class probabilities, with decision_function fallback."""
+        predict_probas = None
+        predict_error = None
+
+        if hasattr(model, "predict_proba"):
+            try:
+                predict_raw = model.predict_proba(model_input)
+                predict_raw = _ensure_numeric_predictions(predict_raw)
+                predict_probas = np.asarray(predict_raw, dtype="float64")
+            except Exception as e:
+                predict_error = e
+
+        if predict_probas is not None:
+            if predict_probas.ndim == 1:
+                if n_labels == 2:
+                    predict_probas = np.column_stack(
+                        [1.0 - predict_probas, predict_probas]
+                    )
+                else:
+                    predict_probas = None
+            elif predict_probas.ndim != 2:
+                predict_probas = None
+
+            if (
+                predict_probas is not None
+                and n_labels is not None
+                and predict_probas.shape[1] != n_labels
+            ):
+                predict_probas = None
+
+        if predict_probas is None and hasattr(model, "decision_function"):
+            scores_raw = model.decision_function(model_input)
+            scores_raw = _ensure_numeric_predictions(scores_raw)
+            predict_probas = self._decision_scores_to_probas(
+                scores_raw, n_labels=n_labels
+            )
+
+        if predict_probas is None:
+            if predict_error is not None:
+                raise ValueError(
+                    "Could not compute class probabilities from model.predict_proba(...)."
+                ) from predict_error
+            raise ValueError(
+                "Could not compute class probabilities: model has neither a working "
+                "predict_proba(...) nor decision_function(...)."
+            )
+
+        if n_labels is not None and predict_probas.shape[1] != n_labels:
+            raise ValueError(
+                f"Expected {n_labels} class probabilities, got shape {predict_probas.shape}."
+            )
+        return predict_probas
+
     @property
     def pred_probas_raw(self):
         """returns pred_probas with probability for each class"""
         if not hasattr(self, "_pred_probas"):
             logger.info("Calculating prediction probabilities...")
-            assert hasattr(
-                self.model, "predict_proba"
-            ), "model does not have a predict_proba method!"
             if self.shap == "skorch":
-                pred_probas_raw = self.model.predict_proba(self.X.values)
-                pred_probas_raw = _ensure_numeric_predictions(pred_probas_raw)
-                self._pred_probas = np.asarray(pred_probas_raw).astype(self.precision)
+                self._pred_probas = self._predict_proba_from_model(
+                    self.model,
+                    self.X.values,
+                    n_labels=len(self.labels),
+                ).astype(self.precision)
             else:
                 warnings.filterwarnings("ignore", category=UserWarning)
-                pred_probas_raw = self.model.predict_proba(self.X)
-                pred_probas_raw = _ensure_numeric_predictions(pred_probas_raw)
-                self._pred_probas = np.asarray(pred_probas_raw).astype(self.precision)
+                self._pred_probas = self._predict_proba_from_model(
+                    self.model,
+                    self.X,
+                    n_labels=len(self.labels),
+                ).astype(self.precision)
                 warnings.filterwarnings("default", category=UserWarning)
         return self._pred_probas
 
@@ -3196,10 +3280,11 @@ def shap_explainer(self):
 
                 def model_predict(data_asarray):
                     data_asframe = pd.DataFrame(data_asarray, columns=self.columns)
-                    pred_probas_raw = self.model.predict_proba(data_asframe)
-                    # Handle XGBoost 3.0+ string predictions (though predict_proba usually returns numeric)
-                    pred_probas_raw = _ensure_numeric_predictions(pred_probas_raw)
-                    return np.asarray(pred_probas_raw)
+                    return self._predict_proba_from_model(
+                        self.model,
+                        data_asframe,
+                        n_labels=len(self.labels),
+                    )
 
                 self._shap_explainer = shap.KernelExplainer(
                     model_predict,
@@ -3750,11 +3835,12 @@ def get_cv_metrics(n_splits):
             ):
                 X_train, X_test = self.X.iloc[train_index], self.X.iloc[test_index]
                 y_train, y_test = self.y.iloc[train_index], self.y.iloc[test_index]
-                preds_raw = (
-                    clone(self.model).fit(X_train, y_train).predict_proba(X_test)
+                fitted_model = clone(self.model).fit(X_train, y_train)
+                preds = self._predict_proba_from_model(
+                    fitted_model,
+                    X_test,
+                    n_labels=len(self.labels),
                 )
-                preds_raw = _ensure_numeric_predictions(preds_raw)
-                preds = np.asarray(preds_raw)
                 for label in range(len(self.labels)):
                     for cut in np.linspace(1, 99, 99, dtype=int):
                         y_true = np.where(y_test == label, 1, 0)
@@ -3981,9 +4067,11 @@ def prediction_result_df(
             else:
                 model_input = sanitize_categorical_predict_input(X_row, self.model)
 
-            pred_probas_raw = self.model.predict_proba(model_input)
-            pred_probas_raw = _ensure_numeric_predictions(pred_probas_raw)
-            pred_probas = np.asarray(pred_probas_raw).squeeze()
+            pred_probas = self._predict_proba_from_model(
+                self.model,
+                model_input,
+                n_labels=len(self.labels),
+            ).squeeze()
             if pred_probas.ndim > 1:
                 pred_probas = pred_probas[0]