Principle:Scikit learn contrib Imbalanced learn Index Balanced Accuracy

Knowledge Sources	On the Effectiveness of Preprocessing Methods When Dealing with Different Levels of Class Imbalance
Domains	Machine_Learning, Model_Evaluation, Imbalanced_Learning
Last Updated	2026-02-09 03:00 GMT

Overview

A meta-metric that adjusts any classification scoring function by weighting it with the dominance relationship between sensitivity and specificity, penalizing classifiers that favor one class over another.

Description

The Index Balanced Accuracy (IBA) wraps an existing metric (like geometric mean or accuracy) and adjusts it based on the dominance factor: the difference between sensitivity and specificity. A classifier that achieves high scores by favoring one class will be penalized. The adjustment formula is:

$IBA=(1+\alpha \times dominance)\times metri{c}^p$

Where dominance = sensitivity - specificity, alpha is a weighting factor, and p = 2 if squared=True (default).

Usage

Use this principle to adjust any metric for class dominance bias. Apply it as a decorator/wrapper around geometric_mean_score or other metrics.

Theoretical Basis

$IBA(\alpha )=(1+\alpha \cdot (sensitivity-specificity))\cdot metri{c}^2$

When the classifier has equal sensitivity and specificity (no dominance), IBA equals the squared metric. When there is class dominance, IBA is adjusted proportionally.

Related Pages

Implemented By

• Implementation:Scikit_learn_contrib_Imbalanced_learn_make_index_balanced_accuracy