FAIR-CARE: A comparative evaluation of unfairness mitigation approaches

Abstract

Bias and unfairness in Machine Learning (ML) are challenging to detect and mitigate, particularly in critical fields such as finance, hiring, and healthcare. While numerous unfairness mitigation techniques exist, most evaluation frameworks assess only a limited set of fairness metrics, primarily focusing on the trade-off between fairness and accuracy. We introduce FAIR-CARE, a new open-source and robust approach that consists of an evaluation pipeline designed for the systematic assessment of unfairness mitigation techniques. Our approach simultaneously evaluates multiple fairness and performance metrics across various ML models. We conduct a comparative analysis on healthcare datasets with diverse distributions—including target class, protected attribute, and their joint distributions—to identify the most effective mitigation technique for each processing type (pre-, in-, and post-processing). Furthermore, we determine the best-performing techniques across different datasets, fairness metrics, performance metrics, and ML models. Finally, we provide practical insights into the application of these techniques, offering actionable guidance for both researchers and practitioners. • Fairness and accuracy can coexist in machine learning models, even on unbalanced datasets. • Mitigation technique performance varies by dataset type and processing stage. • FAIR-CARE identifies top-performing mitigation techniques for each processing type and highlights trade-offs where applicable.

Ähnliche Arbeiten