intrinsic-kappa 0.1.2

Prevalence-independent intrinsic kappa coefficient for classification systems with any number of categories.

intrinsic-kappa

A Python library for computing the prevalence-independent intrinsic kappa coefficient and its one-sided confidence lower bounds for classification systems with any number of categories (NC ≥ 2).

Background

Traditional Cohen's kappa depends on class prevalence — the proportion of instances in each category — which can produce misleading performance estimates when a dataset is imbalanced. The intrinsic kappa solves this by anchoring the chance-agreement term exclusively to the number of categories (1/NC), providing a robust and stable measure of a classifier's inherent discriminatory ability, regardless of sample composition.

This package is based on the methodology published in:

Sanchez-Marquez, R., & Jabaloyes Vivas, J. (2026). A statistical approach to the confusion matrix for classification problems using machine learning. Neurocomputing, 133962. ISSN 0925-2312. https://doi.org/10.1016/j.neucom.2026.133962

Installation

pip install intrinsic-kappa