Robust and Actionable ML via Causality

Abstract: Artificial Intelligence is increasingly deployed in high-stakes fields such as healthcare, where two critical challenges emerge: models must generalize to real-world variations and their predictions must be actionable for decision-makers. The generalization challenge arises when AI is applied to data that differs from its training set, a phenomenon known as distribution shift. The actionability challenge emerges in off-policy evaluation – where we aim to predict the effect of new treatment policies while only having access to historical data collected under different treatment policies. I will present work that tackles these problems, building on the interplay of causality and large-scale machine learning. Drawing on principles from causality, we develop learning rules that induce robustness under distribution shifts. For off-policy evaluation, we introduce a scalable method for causal effect estimation of interventions on when and how to treat. The method is designed to handle treatment timing in observations that are irregularly spread across the timeline, a defining feature of data in healthcare, finance, and other fields. Taken together, these works demonstrate how principled causal approaches can solve fundamental challenges in building reliable AI systems that support decision-making.