关于不确定性——不确定性感知的可解释人工智能系统性调查

This survey focuses on uncertainty-aware explainable artificial intelligence (UAXAI), examining how uncertainty is integrated into explanatory pipelines and the evaluation of these methods. Three main approaches to uncertainty quantification are identified in the literature: Bayesian methods, Monte Carlo methods, and conformal methods. Distinct strategies for incorporating uncertainty into explanations include assessing trustworthiness, constraining models or explanations, and explicitly communicating uncertainty. Evaluation practices are found to be fragmented and largely model-centered, with limited attention to user perspectives and inconsistent reporting of reliability properties such as calibration, coverage, and explanation stability. Recent works have shown a trend towards improving calibration and uncertainty communication.