Unveil Sources of Uncertainty: Feature Contribution to Conformal Prediction Intervals

Uncertainty deserves explanation, not just predictions: Traditional AI explainability methods only explain what a model predicts on average, but ignore how confident the model should be. This paper addresses a critical gap by developing methods to show which input features contribute most to a model's uncertainty—important information for high-stakes decisions where confidence matters as much as the prediction itself.

Game theory provides a rigorous framework for uncertainty attribution: The authors use cooperative game theory (specifically Shapley values and Harsanyi allocations) to fairly distribute responsibility for predictive uncertainty across input features. This mathematical grounding ensures the attribution method is principled and defensible, rather than ad-hoc—crucial for responsible AI systems that must justify their reasoning.

The method works across different types of AI models: By grounding their approach in conformal prediction rather than specific model architectures, the authors created a "model-agnostic" tool that can be applied to any machine learning system. This flexibility increases the practical potential for adoption across diverse AI applications in healthcare, finance, and other sensitive domains.

Computational efficiency was solved to make this practical: The paper proposes a Monte Carlo approximation method with "robust statistical guarantees" that significantly improves runtime, addressing a key barrier to real-world implementation. This means organizations can actually use these uncertainty explanations without prohibitive computational costs.

This toolkit helps communicate AI confidence in high-stakes settings: By systematically explaining which features drive uncertainty, this work provides a structured way for AI systems to communicate not just predictions but also how much to trust them—essential for building appropriate human trust in AI systems used in consequential decisions.