Fairness Evaluation with Item Response Theory

IRT offers a new lens for fairness evaluation: This paper introduces Item Response Theory—a method traditionally used in educational testing—to evaluate fairness in machine learning models. This novel application allows researchers to simultaneously assess model performance and identify which individuals or groups experience unfair treatment.

The framework identifies three key fairness dimensions: Fair-IRT measures a model's "ability to make fair predictions," the "discrimination" of individuals (how differently they're treated across predictions), and the "difficulty" of individuals (how challenging they are for models to predict fairly). This multi-dimensional approach provides more nuanced fairness insights than single-metric evaluations.

Fairness problems can originate from either models or individuals: By analyzing the "flatness" of item characteristic curves, the framework can distinguish whether unfair predictions stem from model bias or from characteristics of specific individuals. This distinction is crucial for determining whether to fix the model or adjust how predictions are interpreted for certain groups.

Real-world applicability across different prediction tasks: The authors demonstrate the framework's effectiveness through case studies in both classification and regression tasks, suggesting it can be practically applied across diverse machine learning applications beyond just binary classification problems.

Enables targeted fairness interventions: By identifying specific individuals or groups that experience discrimination and understanding why, this framework could help practitioners design more targeted fairness improvements rather than applying one-size-fits-all solutions.