Fix macro average calculation in geometric_mean_score for binary classification

imblearn/metrics/_classification.py

@@ -671,16 +671,27 @@ class is unrecognized by the classifier, G-mean resolves to zero. To
     array([0.866...,  0.       ,  0.       ])
     """
     if average is None or average != "multiclass":
-        sen, spe, _ = sensitivity_specificity_support(
+        # LOGIC:
+        # For macro and weighted, calculate per-class first, then average
+        calc_average = None if average in ["macro", "weighted"] else average
+
+        sen, spe, sup = sensitivity_specificity_support(
             y_true,
             y_pred,
             labels=labels,
             pos_label=pos_label,
-            average=average,
+            average=calc_average,
             warn_for=("specificity", "specificity"),
             sample_weight=sample_weight,
         )
 
+        if average in ["macro", "weighted"]:
+            gmean_per_class = np.sqrt(sen * spe)
+            if average == "macro":
+                return np.mean(gmean_per_class)
+            else: # weighted
+                return np.average(gmean_per_class, weights=sup)
+
         return np.sqrt(sen * spe)
     else:
         present_labels = unique_labels(y_true, y_pred)

imblearn/metrics/tests/test_classification.py

@@ -228,9 +228,9 @@ def test_geometric_mean_multiclass(y_true, y_pred, correction, expected_gmean):
 @pytest.mark.parametrize(
     "y_true, y_pred, average, expected_gmean",
     [
-        ([0, 1, 2, 0, 1, 2], [0, 2, 1, 0, 0, 1], "macro", 0.471),
+        ([0, 1, 2, 0, 1, 2], [0, 2, 1, 0, 0, 1], "macro", 0.289),
         ([0, 1, 2, 0, 1, 2], [0, 2, 1, 0, 0, 1], "micro", 0.471),
-        ([0, 1, 2, 0, 1, 2], [0, 2, 1, 0, 0, 1], "weighted", 0.471),
+        ([0, 1, 2, 0, 1, 2], [0, 2, 1, 0, 0, 1], "weighted", 0.289),
         ([0, 1, 2, 0, 1, 2], [0, 2, 1, 0, 0, 1], None, [0.8660254, 0.0, 0.0]),
     ],
 )
@@ -255,7 +255,7 @@ def test_geometric_mean_average(y_true, y_pred, average, expected_gmean):
             [0, 1, 1, 0, 0, 1],
             [1, 2, 1, 1, 2, 1],
             "weighted",
-            0.333,
+            0.236,
         ),
     ],
 )
@@ -277,8 +277,8 @@ def test_geometric_mean_sample_weight(
     [
         ("multiclass", 0.41),
         (None, [0.85, 0.29, 0.7]),
-        ("macro", 0.68),
-        ("weighted", 0.65),
+        ("macro", 0.61),
+        ("weighted", 0.58),
     ],
 )
 def test_geometric_mean_score_prediction(average, expected_gmean):
@@ -547,3 +547,22 @@ def test_macro_averaged_mean_absolute_error_sample_weight():
     )
 
     assert ma_mae_unit_weights == pytest.approx(ma_mae_no_weights)
+
+def test_geometric_mean_score_binary_macro_vs_per_class():
+    """Check that binary + macro average matches mean of per-class scores.
+
+    Regression test for issue where binary macro was computed as 
+    sqrt(mean_tpr * mean_tnr) instead of mean(sqrt(tpr*tnr)).
+    """
+    y_true = [0, 0, 1, 0, 1, 1]
+    y_pred = [0, 0, 0, 0, 0, 1]
+
+    # Calculate per-class scores manually
+    per_class_scores = geometric_mean_score(y_true, y_pred, average=None)
+    expected_macro = np.mean(per_class_scores)
+
+    # Calculate the library's macro score
+    library_macro = geometric_mean_score(y_true, y_pred, average="macro")
+
+    # They should be equal
+    assert library_macro == pytest.approx(expected_macro)