Counterfactual Explanations of Black-box Machine Learning Models using Causal Discovery with Applications to Credit Rating

Causal discovery makes AI explanations more practical: The paper addresses a major limitation in existing explainable AI methods—they typically require knowing the causal relationships between features beforehand. This new framework automatically discovers these relationships from data, making counterfactual explanations feasible for real-world applications where causal structures are unknown.

Counterfactual explanations help users understand "what-if" scenarios: By using counterfactual probabilities, the method can explain predictions by showing what input changes would alter an AI decision. For example, in credit rating decisions, it could show a borrower specifically what factors would need to change to receive a better rating.

Improved accuracy when causal structure is incorporated: Numerical experiments demonstrated that including an estimated causal graph produces more accurate explanatory scores compared to methods that ignore causal relationships, suggesting this approach provides more reliable explanations of AI decisions.

Real-world validation in credit decisions: The framework was tested on actual credit rating data from a Japanese bank, demonstrating practical applicability to high-stakes financial decisions where explainability is critical for regulatory compliance and fairness.

Bridges the gap between black-box models and interpretability: By working with any classification model ("black-box"), this approach makes it possible to explain complex AI systems without requiring simpler, less accurate models—potentially enabling organizations to maintain predictive performance while improving transparency.