FIX OneVsRestClassifier to ensure that predict == argmax(decision_function) (scikit-learn#15504)

Co-authored-by: Lakshmi Krishnan <lakshmik@graphcore.ai>
Co-authored-by: Jérémie du Boisberranger <jeremie@probabl.ai>

Author: 3 people

doc/whats_new/upcoming_changes/sklearn.multiclass/15504.fix.rst

@@ -0,0 +1,3 @@
+- Fix tie-breaking behavior in :class:`multiclass.OneVsRestClassifier` to match
+  `np.argmax` tie-breaking behavior.
+  By :user:`Lakshmi Krishnan <lakrish>`.

sklearn/multiclass.py

@@ -499,10 +499,12 @@ def predict(self, X):
             maxima = np.empty(n_samples, dtype=float)
             maxima.fill(-np.inf)
             argmaxima = np.zeros(n_samples, dtype=int)
-            for i, e in enumerate(self.estimators_):
+            n_classes = len(self.estimators_)
+            # Iterate in reverse order to match np.argmax tie-breaking behavior
+            for i, e in enumerate(reversed(self.estimators_)):
                 pred = _predict_binary(e, X)
                 np.maximum(maxima, pred, out=maxima)
-                argmaxima[maxima == pred] = i
+                argmaxima[maxima == pred] = n_classes - i - 1
             return self.classes_[argmaxima]
         else:
             thresh = _threshold_for_binary_predict(self.estimators_[0])

sklearn/tests/test_multiclass.py

@@ -82,6 +82,25 @@ def test_check_classification_targets():
         check_classification_targets(y)
 
 
+def test_ovr_ties():
+    """Check that ties-breaking matches np.argmax behavior
+
+    Non-regression test for issue #14124
+    """
+
+    class Dummy(BaseEstimator):
+        def fit(self, X, y):
+            return self
+
+        def decision_function(self, X):
+            return np.zeros(len(X))
+
+    X = np.array([[0], [0], [0], [0]])
+    y = np.array([0, 1, 2, 3])
+    clf = OneVsRestClassifier(Dummy()).fit(X, y)
+    assert_array_equal(clf.predict(X), np.argmax(clf.decision_function(X), axis=1))
+
+
 def test_ovr_fit_predict():
     # A classifier which implements decision_function.
     ovr = OneVsRestClassifier(LinearSVC(random_state=0))