Quotient DAGs for Off-Policy Evaluation:Forward-Flow Importance Sampling and Exact Slate Propensities
Two new research papers explore advanced techniques for off-policy evaluation (OPE) in machine learning, a critical process for assessing the performance of new policies using existing data. The first paper introduces "Quotient DAGs" to handle situations where the reward depends only on an unordered set of items, even if the generation process is ordered, thereby reducing nuisance variance. The second paper, "CANDOR," proposes a doubly robust OPE estimator that effectively leverages imperfect expert-annotated counterfactual samples, particularly for healthcare applications, by incorporating annotations into the reward model component. AI
IMPACT These papers introduce novel methods for off-policy evaluation, potentially improving the reliability and safety of deploying new policies in critical domains like healthcare.