Constraint Preservation in Counterfactual Explanations for Credit Scoring

Sammanfattning

The increasing adoption of complex machine learning models in high-stakes financial applications has intensified the demand for transparency and accountability in decision making. Explainable Artificial Intelligence (XAI) provides a range of approaches to address these concerns, with counterfactual (CF) explanations offering intuitive, actionable insights by illustrating minimal changes required to alter a model’s prediction. However, conventional CF generation methods often neglect inherent relationships between features, leading to infeasible or inconsistent scenarios that undermine their practical utility. This thesis investigates the integration of precise mathematical and predictive feature dependencies into the CF generation process to preserve domain-specific constraints in financial risk assessment. Building on the Diverse Counterfactual Explanations (DiCE) framework, a constraint-aware extension is implemented to enforce exact relationships during CF generation, ensuring that generated scenarios remain realistic and computationally consistent. The approach is evaluated using financial datasets for credit scoring, examining the impact of relationship preservation on explanation quality, feature modification patterns, and operational performance. Results demonstrate that constraint-aware CF explanations improve plausibility and alignment with domain knowledge, thereby enhancing their value as decision-support tools in finance. The findings start to bridge the gap between technically sound explanations and those that deliver actionable, context-aware guidance in real-world applications.