Status: accepted · Date: 2026-08-14

balanced_accuracy_score is the mean recall over the classes that occur at least once in y_true — not over every class a caller declared. A class that is predicted but never true has an undefined recall (0/0) and scikit-learn drops it from the mean rather than reading it as zero: on y_true=[0,0,1], y_pred=[0,2,1], class 2 is predicted but never true, and mean(recall_0=0.5, recall_1=1.0) = 0.75, not mean(0.5, 1.0, 0) = 0.5. adjusted=True rescales that mean by how many classes were kept, not by how many were declared: (score - 1/kept) / (1 - 1/kept).

With exactly one class kept, 1/kept is 1, so the rescale's denominator is 1 - 1 = 0. What that division returns then depends only on the numerator, which IEEE 754 already defines: 0.0 / 0.0 is NaN, and a negative numerator over zero is -Infinity. Measured against the oracle venv (scikit-learn 1.9.0): balanced_accuracy_score([1,1], [1,1], adjusted=True) is nan (the one kept class's recall is exactly 1.0, so the numerator is 0.0), and balanced_accuracy_score([0,0], [0,1], adjusted=True) is -inf (recall 0.5, numerator -0.5).

BalancedAccuracy.Score does not special-case the single-kept-class edge. chance = 1.0 / kept and (score - chance) / (1.0 - chance) are computed exactly as written; when kept == 1 the denominator is 0.0 and .NET's own IEEE 754 division produces the same NaN/-Infinity split scikit-learn does, with no branch needed to reproduce it.

• The <remarks> on BalancedAccuracy.Score(ConfusionMatrix, bool) carries a pointer here instead of restating the averaging rule and the edge case.
• Verified by BalancedAccuracyTests.Adjusted_divides_by_zero_when_a_single_class_is_kept, which asserts double.IsNaN and double.IsNegativeInfinity for the two cases above, and by the Matches_sklearn_with_adjusted oracle theory, whose corpus (tests/oracles/classification_metrics.json) carries further "NaN" fixtures that this same code path produces with no dedicated branch.
• The "average runs over kept classes, not declared ones" rule itself is pinned separately by BalancedAccuracyTests.Averages_over_the_classes_it_kept_not_over_all_of_them and Adjusted_divides_by_the_classes_it_kept.