Comparables XAI: Faithful Example-based AI Explanations with Counterfactual Trace Adjustments

Example-based explanations can be made more faithful by tracing how individual features change: The paper demonstrates that showing users step-by-step how each attribute of a comparison example differs from the decision being explained—rather than presenting all differences at once—significantly improves how accurately people understand AI decisions.

Real estate valuation methods offer a practical model for AI transparency: By adapting the "Comparables" approach used by real estate appraisers (comparing properties with known values and adjusting for specific differences), the researchers created a framework that users found more intuitive and trustworthy than traditional explanation methods.

User studies confirm this approach reduces uncertainty in understanding: The research included human testing showing that people had "narrower uncertainty bounds" when using Trace-adjusted Comparables compared to other explanation methods, meaning they felt more confident and precise about their understanding of why the AI made a particular decision.

This method outperformed standard technical approaches: The Comparables XAI approach achieved better results on multiple measures—including "highest XAI faithfulness and precision" and "user accuracy"—compared to linear regression explanations and other comparison-based methods, suggesting it's a more reliable way to explain AI decisions to non-technical users.