Comparative Analysis of Pre-Processing, Inprocessing and Post-Processing Methods for Bias Mitigation: A Case Study on Adult Dataset

This research performs deep comparative analysis on the three major bias mitigation strategies, namely, pre-processing, in-processing, and post-processing, with the adult dataset to examine and solve bias issues in income prediction models. The study compares how each approach performs on the reduction of algorithmic bias while keeping the overall performance of the model in check. The 28 percentage points of demographic disparity reduction achieved from the pre-processing approach reflect its contribution in handling data-representation-rooted biases and utility in managing demographic disparities. In the modification of the model training process to reduce discrimination, the inprocessing technique exhibited a 31 percentpoint improvement in equal opportunity while equitably ensuring favourable outcome through demographic groups. The post-processing method refers to modifying the model's predictions after training, and it has achieved up to 25 percent improvements over the intermediate results within the fairness metrics. This would make bias mitigation accessible as an option for such existing models. Indeed, this broader comparison provides the most critical insights on what each approach is particularly good or bad at, thus supporting practical selection of bias mitigation strategies aligned with specific fairness and accuracy goals. The results herein contribute another facet to the growing body of work in fair machine learning: namely the practical trade-offs incurred in real-world predictive modelling applications to achieve fairness.