Big data has enabled decision-makers to personalize decisions in domains like personalized medicine and online advertising. We formulate this problem as a multi-armed bandit with covariates, and present two papers that address practical challenges in these settings.

First, user covariates are often high-dimensional. However, existing methods become impractical in the presence of such big data, since they require the decision-maker to experiment on a very large number of initial users. We present a new efficient algorithm based on the LASSO estimator, and establish that it achieves near-optimal performance; the key step is proving a statistical convergence guarantee for the LASSO estimator despite the non-i.i.d. data induced by the bandit policy. We evaluate our algorithm on the real-world problem of warfarin dosing, and find that it outperforms existing bandit methods as well as physicians to correctly dose a majority of patients.

Second, bandit methods rely on occasional experimentation to balance an exploration-exploitation tradeoff. However, experimentation may be costly or unethical in practice. We prove that experimentation is unnecessary under certain assumptions, particularly if user covariates are sufficiently diverse. We then present a new algorithm that uses observed data to determine whether experimentation is necessary. This algorithm significantly reduces experimentation in simulations.
 

Biography
Hamsa is a fifth-year PhD student in Stanford's Electrical Engineering department and is advised by Mohsen Bayati in the GSB.  She graduated summa cum laude from Harvard in 2012 with an A.M. in theoretical physics, and a A.B. in physics and mathematics.  Her primary research interests center around (a) optimizing service operations by developing novel data-driven statistical decision-making tools using techniques from machine learning, and (b) designing improved performance-based contracts using detailed outcomes data on strategic firms and workers. She is particularly interested in healthcare applications where cost and quality pose serious concerns, and the growing availability of staff and patient data offers an opportunity to significantly improve outcomes through data-driven methods.