Primary Submission Category: Machine Learning and Causal Inference
Personalization to One of Many Arms
Authors: Rahul Ladhania, Jann Spiess, Lyle Ungar,
Presenting Author: Rahul Ladhania*
We consider learning personalized assignments among potentially many treatment arms from a randomized controlled trial. In a theoretical model, we illustrate how a high number of treatment arms makes finding the best arm hard, while we can still achieve sizable welfare gains from personalization by direct optimization. In a practical implementation, we propose methods that optimize treatment assignment specifically in the case of many treatment arms. First, we consider a regularized forest-based assignment algorithm based on greedy recursive partitioning that includes shrinkage across treatment arms. Second, we propose a clustering scheme that combines treatment arms with consistently similar outcomes. In a simulation study, we compare the performance of these approaches to predicting arm-wise outcomes separately, and document gains of directly optimizing the treatment assignment and including regularization and clustering in the underlying model construction.