scikit-learn · jeremiedbb · May 7, 2025 · Apr 20, 2025 · Apr 22, 2025 · Apr 22, 2025
diff --git a/doc/whats_new/upcoming_changes/sklearn.multiclass/31228.fix.rst b/doc/whats_new/upcoming_changes/sklearn.multiclass/31228.fix.rst
@@ -0,0 +1,5 @@
+- The `predict_proba` method of :class:`sklearn.multiclass.OneVsRestClassifier` now
+  returns zero for all classes when all inner estimators never predict their positive
+  class.
+  By :user:`Luis M. B. Varona <Luis-Varona>`, :user:`Marc Bresson <MarcBresson>`, and
+  :user:`Jérémie du Boisberranger <jeremiedbb>`.
diff --git a/sklearn/multiclass.py b/sklearn/multiclass.py
@@ -553,8 +553,10 @@ def predict_proba(self, X):
             Y = np.concatenate(((1 - Y), Y), axis=1)
 
         if not self.multilabel_:
-            # Then, probabilities should be normalized to 1.
-            Y /= np.sum(Y, axis=1)[:, np.newaxis]
+            # Then, (nonzero) sample probability distributions should be normalized.
+            row_sums = np.sum(Y, axis=1)[:, np.newaxis]
+            np.divide(Y, row_sums, out=Y, where=row_sums != 0)
+
         return Y
 
     @available_if(_estimators_has("decision_function"))

diff --git a/sklearn/tests/test_multiclass.py b/sklearn/tests/test_multiclass.py
@@ -6,6 +6,7 @@
 from numpy.testing import assert_allclose
 
 from sklearn import datasets, svm
+from sklearn.base import BaseEstimator, ClassifierMixin
 from sklearn.datasets import load_breast_cancer
 from sklearn.exceptions import NotFittedError
 from sklearn.impute import SimpleImputer
@@ -429,6 +430,31 @@ def test_ovr_single_label_predict_proba():
     assert not (pred - Y_pred).any()
 
 
+def test_ovr_single_label_predict_proba_zero():
+    """Check that predic_proba returns all zeros when the base estimator
+    never predicts the positive class.
+    """
+
+    class NaiveBinaryClassifier(BaseEstimator, ClassifierMixin):
+        def fit(self, X, y):
+            self.classes_ = np.unique(y)
+            return self
+
+        def predict_proba(self, X):
+            proba = np.ones((len(X), 2))
+            # Probability of being the positive class is always 0
+            proba[:, 1] = 0
+            return proba
+
+    base_clf = NaiveBinaryClassifier()
+    X, y = iris.data, iris.target  # Three-class problem with 150 samples
+
+    clf = OneVsRestClassifier(base_clf).fit(X, y)
+    y_proba = clf.predict_proba(X)
+
+    assert_allclose(y_proba, 0.0)
+
+
 def test_ovr_multilabel_decision_function():
     X, Y = datasets.make_multilabel_classification(
         n_samples=100,