Impact of Bias Mitigation on Credit Scoring

This research article examines the impact of bias mitigation techniques on credit scoring systems, focusing on the balance between performance, fairness, and explainability. It outlines the necessity for machine learning models in high-stakes decision-making to not only be accurate but also to ensure fair treatment and provide explainability, as mandated by regulatory frameworks such as the EU's AI Act and the GDPR. The paper discusses counterfactual explanations as a method to meet these requirements, illustrating how minor changes in input can affect model outcomes. It details various bias mitigation techniques that alter training data, classifiers, or prediction probabilities, and their implications for explanation methods. The analysis includes multiple bias mitigation techniques and real-world datasets, revealing that no single method performs best across all dimensions. Some methods enhance fairness at the expense of accuracy, while others improve predictive performance but may compromise fairness. The findings emphasize the importance of joint evaluation frameworks that integrate predictive performance, fairness, and explainability in AI systems.