Informal Research Talk Angela Zhou from Cornell Tuesday, April 30 - 12:30 pm Friend Center - Room 112 Host: Barbara Engelhardt Safely Learning to Personalize from Observational Data Some of the most impactful applications of machine learning are not just about prediction but are rather about taking the right action, directed at the right target, at the right time. The goal of our work is to extract causal-effect-maximizing personalized decision rules from observational data in sensitive applications. Observational data, the norm rather than the exception in domains such as medicine and civics, lack experimental manipulation so that isolated causal effects are obscured by complex selection processes, a phenomenon known as confounding. We develop a framework for learning personalized decision policies from observational data while accounting for possible unobserved confounding in the data-generating process. In contrast, the previous literature assumes that all confounders are observed, an unverified assumption that is generally violated in practice. Under the Neyman-Rubin potential outcomes framework, policy evaluation is analogous to assessing subgroup treatment effects: policy learning parametrizes the subgroup by a machine learning predictor of who should be treated. We calibrate policy learning for realistic violations of this unverifiable assumption with uncertainty sets for inverse propensity weights motivated by sensitivity analysis in causal inference. Our framework for confounding-robust policy improvement optimizes the minimax regret of a candidate policy against a baseline standard-of-care policy over an uncertainty set. We prove that if the uncertainty set is well-specified, our robust policy, when applied in practice, will do no worse than the baseline and improve upon it if possible, and that it is asymptotically minimax optimal. We use efficient algorithmic solutions to optimize over parametrized spaces of decision policies such as logistic treatment assignment and decision trees. We assess our methods on synthetic data and on a large clinical trial, demonstrating that hidden confounding can hinder existing policy learning approaches, while our robust approach guarantees safety and focuses on well-evidenced improvement, a necessity for making personalized treatment policies learned from observational data reliable in practice. A conference version of this work appeared at Neurips 2018 and an extended version is under Major Revision at Management Science. Bio: Angela Zhou is a third-year PhD student at Cornell in Operations Research and Information Engineering. She is currently located at Cornell Tech (NYC) where she is very fortunate to be advised by Nathan Kallus. Her work is supported on a NDSEG fellowship. Previously, she studied Operations Research and Financial Engineering at Princeton University. -- Jonathan Lu | Princeton University | Class of 2019 Department of Computer Science, Master's Program jhlu@princeton.edu | (510) 779-4158