FIX sensitivity_specificity_support: correct specificity with sample_weight (#1180)

[immu4989]

Reference Issues/PRs

Fixes #1180.

What does this implement/fix? Explain your changes.

sensitivity_specificity_support computes an incorrect specificity when sample_weight is provided — the value can even exceed 1, which is impossible for a rate. Because specificity_score, geometric_mean_score, classification_report_imbalanced and the make_index_balanced_accuracy wrappers all delegate to it, the bug propagates through the whole metrics module.

Root cause. When sample_weight is given, tp_sum, pred_sum and true_sum are weighted sums, but the true-negative count was formed using the raw sample count:

tn_sum = y_true.size - (pred_sum + true_sum - tp_sum)

Mixing a count (y_true.size) with weighted sums makes tn_sum wrong — it can go negative — and the downstream specificity = tn_sum / (tn_sum + pred_sum - tp_sum) is then wrong as well.

import numpy as np
from imblearn.metrics import sensitivity_specificity_support

_, spe, _ = sensitivity_specificity_support(
    [0, 0, 1, 1], [0, 1, 1, 1], sample_weight=[1.0, 2.0, 3.0, 4.0], average=None
)
print(spe)
# master: [1.         1.66666667]   <- 1.667 is impossible
# fixed:  [1.         0.33333333]   <- TN=1, FP=2 -> 1/(1+2)

Fix. Use the total sample weight as the population size when sample_weight is provided, falling back to y_true.size otherwise. The unweighted path is unchanged (every term scales uniformly there, so it was already correct).

Scope: a module-wide audit, single root cause

I audited every sample_weight-aware metric in imblearn.metrics using the invariant that integer sample_weight must equal physically repeating each sample. The results:

Metric	Before fix	After fix
sensitivity_specificity_support (sensitivity)	✅ correct	✅ correct
sensitivity_specificity_support (specificity)	❌ wrong (negative TN, values > 1)	✅ correct
specificity_score	❌ wrong (delegates)	✅ correct
geometric_mean_score	❌ wrong, incl. nan (delegates)	✅ correct
make_index_balanced_accuracy(...) wrappers	❌ wrong (delegates)	✅ correct
classification_report_imbalanced	❌ wrong specificity column (delegates)	✅ correct
macro_averaged_mean_absolute_error	✅ correct (independent path)	✅ correct

So the entire weighted-metric bug class traces to this one line; there are no other independent weighting bugs in the module.

Tests

• The existing test_geometric_mean_sample_weight parametrization asserted the weighted value produced under the buggy behaviour (0.333). I corrected it to the right value (0.609), verified by an independent hand-computation of the weighted confusion matrix (per-class specificity [0.667, 0.5], sensitivity [1.0, 0.5]).
• Added test_sensitivity_specificity_support_sample_weight (the [BUG] sensitivity_specificity_support returns wrong specificity with sample_weight (can exceed 1) #1180 reproducer; asserts specificity ≤ 1; integer weights == repeated samples).
• Added a parametrized test_metrics_sample_weight_repeat_equivalence that enforces the repeat-equivalence invariant across the whole delegation chain (sensitivity, specificity, specificity_score, geometric_mean_score, the IBA wrappers) and every average mode (None/macro/weighted/micro), plus a [0, 1] rate-bound check — 28 cases. This locks the fix in across the module, not just where the bug originated.
• Full metrics suite passes locally: pytest imblearn/metrics/ → 236 passed (scikit-learn 1.9.0).