Counterfactual Inference with Unobserved Confounding via Exponential Family

Watch a 44-minute lecture from Harvard CMSA featuring MIT professor Devavrat Shah exploring counterfactual inference with unobserved confounding through exponential family modeling. Learn about personalized decision-making challenges in recommender systems, focusing on how to infer user engagement with different recommendation sequences while accounting for unobserved factors. Discover a computationally efficient method for learning distribution parameters with estimation error scaling linearly with metric entropy. Explore sufficient conditions for compactly supported distributions satisfying logarithmic Sobolev inequality, and understand the application of these concepts in sequential recommender systems, measurement error imputation, and undirected graphical models. The lecture covers theoretical foundations including maximum likelihood estimation, proper loss functions, and parameter estimation techniques for handling heterogeneous users with single trajectory observations.