Generating Counterfactual and Contrastive Explanations using SHAP

SHAP offers a practical tool for legal compliance: The paper demonstrates that SHAP (Shapley Additive Explanations) can be used to generate explanations required under GDPR and similar regulations, making it a concrete method for organizations to meet legal requirements for model transparency.

Two explanation types serve different purposes: The research identifies contrastive explanations (showing why one prediction was made instead of another) and counterfactual explanations (showing what would need to change for a different outcome) as particularly suited to human understanding, offering complementary ways to communicate AI decisions to non-experts.

Model-agnostic approach increases accessibility: By proposing a method that works across different types of AI models regardless of their underlying architecture, the paper suggests that organizations don't need to rebuild systems to achieve explainability—existing models can be made interpretable after deployment.

Practical implementation is feasible: Testing the approach on standard datasets (IRIS, Wine Quality, and Mobile Features) demonstrates that generating these explanations is technically achievable in real-world scenarios, not just theoretical.

Counterfactual explanations enable actionable insights: By generating counterfactual datapoints, the method allows stakeholders to understand specific, concrete changes that would alter an AI system's decision, which is particularly valuable for individuals affected by automated decisions who want to know how to achieve different outcomes.