Inverse Propensity

Inverse propensity scoring (IPS) is a crucial technique in causal inference, particularly for estimating treatment effects in observational studies and evaluating policies using offline data (off-policy evaluation). Current research focuses on mitigating the high variance and bias inherent in IPS estimators, especially in high-dimensional settings, through methods like data-dependent coarsening, proximal policy optimization, and marginalized IPS (MIPS) which leverage action embeddings. These advancements improve the robustness and accuracy of causal effect estimations and policy evaluations across diverse applications, including recommendation systems, ranking algorithms, and contextual bandits, leading to more reliable and impactful results in these fields.