MISSING SOMETHING? ACCOUNTING FOR INPUT QUALITY UNCERTAINTY TO IMPROVE AI CONFIDENCE SCORES

Abstract: Current methods for the assessment of the uncertainty associated with AI predictions fail to account for input quality uncertainty associated with missing data in inference instances. Thus, AI systems may provide overconfident predictions that do not reflect the actual uncertainty, leading to potential misjudgments in decision-making processes in human-AI teams. In this paper, we propose a novel approach to account for input quality uncertainty originating from missing values in the assessment of confidence scores to ensure a well-founded and transparent consideration and representation of uncertainty in the context of AI-supported decision-making. We evaluate our approach based on four datasets against competing state-of-the-art approaches. Results demonstrate improved confidence calibration and robustness to varying proportions of missing data, without compromising on prediction accuracy. By providing better calibrated and thus trustworthy uncertainty estimates, our approach facilitates appropriate reliance on AI systems, improving human-AI team performance in critical decision-making contexts.