Realistic Counterfactual Explanations via Denial Constraints

Counterfactual explanations can be unrealistic: Current AI explainability methods often generate counterfactual explanations (small changes to data that flip a model's prediction) that don't reflect real-world possibilities, limiting their practical usefulness for understanding why a model made a decision.

Realism matters for trustworthy AI: The paper demonstrates that explanations based on realistic scenarios are more valuable than those based on hypothetical but impossible situations—for example, explaining a loan denial by suggesting someone change their age, which is not actionable or realistic.

Denial constraints capture real-world rules: The researchers use "denial constraints" (logical rules derived from actual datasets) to ensure counterfactual explanations respect patterns that exist in real data, making explanations more grounded and applicable to actual decision-making contexts.

Realism can be achieved with minimal trade-offs: Experiments across four datasets show that the method successfully generates realistic counterfactual explanations while maintaining relatively small distances from original instances and preserving diversity of explanations—meaning users don't have to sacrifice explanation quality for realism.

Bridging AI and data management improves explainability: By combining techniques from explainable AI with data cleaning methods from database management, the paper shows that responsible AI explanations benefit from interdisciplinary approaches that ground algorithmic outputs in real-world data patterns.