Concisely Explaining the Doubt: Minimum-Size Abductive Explanations for Linear Models with a Reject Option

Reject options are critical for trustworthy AI in high-stakes domains: The paper emphasizes that in healthcare and finance, AI systems must be able to abstain from making decisions when evidence is insufficient, rather than forcing a prediction. This capability is essential for supporting informed human intervention in cases of uncertainty.

Explanations for rejections are as important as explanations for decisions: The research highlights that current AI explainability work focuses mainly on explaining what a model decides, but neglecting to explain why a model rejects a case leaves humans without the information needed to make informed decisions. Clear rejection explanations are necessary for trustworthiness.

Fidelity and efficiency matter equally for practical AI systems: The paper identifies a key tension in real-world deployment: explanations must be both faithful to how the model actually works and fast enough for real-time use. Previous approaches sacrificed one for the other, but this work aims to achieve both simultaneously for linear models.

Minimum-size explanations are more useful than just any explanation: Rather than providing lengthy lists of factors influencing a decision or rejection, the research focuses on finding the smallest set of essential reasons. This conciseness makes explanations more interpretable and actionable for human decision-makers.

Technical advances enable previously intractable problems: By adapting existing algorithms, this work makes it computationally feasible to generate optimal explanations for rejection cases in linear models—a problem that was previously either unsolved or only partially addressed in restricted settings.