Evaluating Bias Mitigation Techniques in Credit and Marketing Models: Balancing Fairness and Performance

Algorithmic decision making in financial services often amplifies existing societal biases due to imbalanced data and historical discrimination. Ensuring fairness in machine learning models, particularly within credit scoring and marketing domains, is therefore both an ethical and regulatory imperative. This paper presents a comprehensive empirical evaluation of prominent bias mitigation techniques using both pre-processing and post-processing methods from Fairlearn and AIF360 frameworks. Using four benchmark datasets Synthetic, German Credit, Bank Marketing, and Credit Card Default, we analyze fairness across protected attributes such as gender, age, and marital status. Models including Logistic Regression (LR) and Random Forest (RF) serve as baselines, while bias mitigation is applied using Exponentiated Gradient Reduction, Threshold Optimization, Reweighing, and Equalized Odds Postprocessing. Performance is evaluated across metrics including AUC, Accuracy, Disparate Impact (DI), Demographic Parity (DP_diff), and Equalized Odds (EO_diff) differences. Results show that while mitigation methods consistently reduce bias metrics (DP_diff, EO_diff, DI) across all datasets, they incur a minimal performance cost (average AUC drop less than 1.5%). AIF360 Reweighing and Fairlearn Threshold Optimizer are shown to achieve the best overall fairness–performance balance, with method effectiveness being highly dependent on the type of inherent data bias. The findings highlight the importance of contextual bias measurement and dataset specific fairness strategies in responsible AI deployment for financial decision making.