The work “A Causal Framework for Evaluating Deferring Systems” (link to camera-ready version will be added asap) with Filippo Palomba, José M. Alvarez and Salvatore Ruggieri has been accepted at the The 28th International Conference on Artificial Intelligence and Statistics (AISTATS).
This paper proposes a novel framework for evaluating deferring systems, which are ML models that can defer difficult instances to human experts. We show that the problem of learning to defer can be naturally framed as a causal inference problem using the potential outcomes framework. We provide theoretical results and showcase how to reason on causal effects in our experimental evaluation.
A final thanks to all my amazing co-authors and see you in Mao Khao 🏝️🍹⛱️🌞🌊.