A Comparison of Methods for Treatment Assignment with an Application to Playlist Generation

This study presents a systematic comparison of methods for individual treatment assignment. We group the various proposed in literature into three general classes algorithms (or metalearners): learning models to predict outcomes (the O-learner), causal effects E-learner), and optimal assignments A-learner). discuss how metalearners differ their level generality objective function, which has critical implications modeling decision making. Notably, we demonstrate that optimizing prediction or is not same as assignments, suggesting that, general, A-learner should lead better than other metalearners. then compare context choosing, each user, best algorithm playlist generation order optimize engagement. first different on real-world application at scale (based more half billion assignments). In addition supporting our analytical findings, results show large A/B tests can provide substantial value treatment-assignment policies, rather simply choosing variant performs average.