General procedure to measure fairness in regression problems

Abstract Fairness in artificial intelligence has emerged as a critical ethical concern, with most research focusing on classification tasks despite the prevalence of regression problems in real-world applications. We address this gap by presenting a general procedure for measuring fairness in regression problems, focusing on statistical parity as a fairness metric. Through extensive experimental analysis, we evaluate how different methodological choices, such as discretization methods, algorithm selection, and parameter optimization, impact fairness outcomes in regression tasks. Our primary contribution is a systematic framework that helps practitioners assess and compare fairness across various approaches to solving regression problems, providing clear guidelines for selecting appropriate strategies based on specific problem requirements. The results demonstrate the importance of carefully considering procedural decisions when evaluating fairness in regression contexts, as these choices influence both model performance and fairness outcomes.