Calibrated Explanations: with Uncertainty Information and Counterfactuals

Artificial Intelligence (AI) has become an integral part of decision support systems (DSSs) in various domains, but the lack of transparency in the predictive models used in AI-based DSSs can lead to misuse or disuse. Explainable Artificial Intelligence (XAI) aims to create AI systems that can explain their rationale to human users. Local explanations in XAI can provide information about the causes of individual predictions in terms of feature importance, but they suffer from drawbacks such as instability. To address these issues, we propose a new feature importance explanation method, Calibrated Explanations (CE), which is based on Venn-Abers and calibrates the underlying model while generating feature importance explanations. CE provides fast, reliable, stable, and robust explanations, along with uncertainty quantification of the probability estimates and feature importance weights. Furthermore, the method is model agnostic with easily understood conditional rules and can also generate counterfactual explanations with uncertainty quantification.