A Bayesian explanation of machine learning models based on modes and functional ANOVA

The paper addresses a gap in explainability by focusing on why predictions deviate from expected outcomes rather than just explaining individual predictions. Most AI explanation methods describe what features matter for a specific prediction, but this work instead asks: given an unexpected result, which features caused this deviation? This "inverse" approach may be more useful for identifying when and why AI systems fail.

The method uses Bayesian statistics to identify the "true" influential features rather than relying on average values. The authors argue their approach is more robust than SHAP values (a popular explanation method) because it conditions on the actual observed outcome rather than comparing to mean values, potentially providing more accurate explanations of individual cases.

The explanation method ranks features by their "distance" in functional ANOVA decomposition, which the authors claim is more intuitive for humans to understand. Rather than abstract numerical scores, this approach may help non-technical stakeholders better grasp which factors actually drove an AI decision.

The computational cost doesn't increase with the number of features in the dataset. This "dimension-independent" efficiency means the method could scale to real-world AI systems with hundreds or thousands of potential features, making it practically viable for responsible AI auditing.

This work contributes to responsible AI by enabling better investigation of unexpected or problematic predictions, which is crucial for detecting bias, errors, and ensuring AI systems behave as intended before deployment.