Abstract

Machine Learning (ML) algorithms play a vital role in decision-making across multiple sectors such as healthcare, finance, education, and law enforcement. Despite their widespread use, concerns about fairness and bias within these systems present critical ethical and social challenges. To address these issues, three major fairness libraries-Microsoft’s Fair-learn, IBM’s AIF360, and Google’s What-If Tool-were utilized. These frameworks offer comprehensive capabilities for assessing fairness, visualizing outcomes, and implementing bias mitigation techniques. This study focuses on evaluating and mitigating bias in a structured dataset using the CatBoost Classifier model. Its primary objective is to compare the performance of mitigation algorithms from two fairness libraries, applying them individually across the three stages of the ML lifecycle (pre-processing, in-processing, and post-processing), as well as sequentially across multiple stages. The results indicate that certain sequential applications improve bias reduction while preserving model performance. A publicly available dataset from Kaggle was used for experimentation, providing a realistic context for examining fairness within machine learning workflows.

Keywords:Machine learning fairness; Bias analysis