A Comparative Analysis of Bias Mitigation Methods in Machine Learning

Fairness in AI and Machine Learning is gaining traction everyday as more people are understanding the profound effects it can have on a model's predictions. While there is scope for advancement in this field, multiple bias mitigation methods and fairness metrics have already been developed. This study aims to conduct a comparative analysis of different bias mitigation techniques using the UCI Adult Income dataset while deploying Logistic Regression as the base model. This study focuses solely on analyzing technical methods to mitigate bias, excluding non-technical approaches. One of pre-processing, in-processing, and post-processing methods are chosen and implemented for analysis using Python and the AI Fairness 360 (AIF360) toolkit. We evaluate the model's predictions before and after debiasing, and we use fairness and classification metrics to compare the impact of different bias mitigation strategies. Additionally, statistical tests are utilized to determine if the methods make any significant differences.