Preventing Failures Due to Dataset Shift: Learning Predictive Models That Transport

Classical supervised learning produces unreliable models when training and distributions differ, with most existing solutions requiring samples the target domain. We propose a proactive approach which learns a in the training domain that will generalize to the target domain incorporating prior knowledge of aspects of the data generating process that expected to differ as expressed in a causal selection diagram. , we remove variables generated by unstable mechanisms from the factorization to yield the Surgery Estimator---an interventional that is invariant to the differences across environments. We prove the surgery estimator finds stable relationships in strictly more than previous approaches which only consider conditional, and demonstrate this in simulated experiments. We also evaluate real world data for which the true causal diagram is unknown, performing against entirely data-driven approaches.