Policy Learning with Adaptively Collected Data

Learning optimal policies from historical data enables the gains from personalization to be realized in a wide variety of applications. The growing policy learning literature focuses on a setting where the treatment assignment policy does not adapt to the data. However, adaptive data collection is becoming more common in practice, from two primary sources: 1) data collected from adaptive experiments that are designed to improve inferential efficiency; 2) data collected from production systems that are adaptively evolving an operational policy to improve performance over time (e.g. contextual bandits). In this paper, we aim to address the challenge of learning the optimal policy with adaptively collected data and provide one of the first theoretical inquiries into this problem. We propose an algorithm based on generalized augmented inverse propensity weighted estimators and establish its finite-sample regret bound. We complement this regret upper bound with a lower bound that characterizes the fundamental difficulty of policy learning with adaptive data. Finally, we demonstrate our algorithms effectiveness using both synthetic data and public benchmark datasets.

Online Multi-Armed Bandits with Adaptive Inference

During online decision making in Multi-Armed Bandits (MAB), one needs to conduct inference on the true mean reward of each arm based on data collected so far at each step. However, since the arms are adaptively selected--thereby yielding non-iid data--conducting inference accurately is not straightforward. In particular, sample averaging, which is used in the family of UCB and Thompson sampling (TS) algorithms, does not provide a good choice as it suffers from bias and a lack of good statistical properties (e.g. asymptotic normality). Our thesis in this paper is that more sophisticated inference schemes that take into account the adaptive nature of the sequentially collected data can unlock further performance gains, even though both UCB and TS type algorithms are optimal in the worst case. In particular, we propose a variant of TS-style algorithms--which we call doubly adaptive TS--that leverages recent advances in causal inference and adaptively reweights the terms of a doubly robust estimator on the true mean reward of each arm. Through 20 synthetic domain experiments and a semi-synthetic experiment based on data from an A/B test of a web service, we demonstrate that using an adaptive inferential scheme (while still retaining the exploration efficacy of TS) provides clear benefits in online decision making: the proposed DATS algorithm has superior empirical performance to existing baselines (UCB and TS) in terms of regret and sample complexity in identifying the best arm. In addition, we also provide a finite-time regret bound of doubly adaptive TS that matches (up to log factors) those of UCB and TS algorithms, thereby establishing that its improved practical benefits do not come at the expense of worst-case suboptimality.

Derandomizing Knockoffs

Model-X knockoffs is a general procedure that can leverage any feature importance measure to produce a variable selection algorithm, which discovers true effects while rigorously controlling the number or fraction of false positives. Model-X knockoffs is a randomized procedure which relies on the one-time construction of synthetic (random) variables. This paper introduces a derandomization method by aggregating the selection results across multiple runs of the knockoffs algorithm. The derandomization step is designed to be flexible and can be adapted to any variable selection base procedure to yield stable decisions without compromising statistical power. When applied to the base procedure of Janson et al. (2016), we prove that derandomized knockoffs controls both the per family error rate (PFER) and the k family-wise error rate (k-FWER). Further, we carry out extensive numerical studies demonstrating tight type-I error control and markedly enhanced power when compared with alternative variable selection algorithms. Finally, we apply our approach to multi-stage genome-wide association studies of prostate cancer and report locations on the genome that are significantly associated with the disease. When cross-referenced with other studies, we find that the reported associations have been replicated.

Dynamic Batch Learning in High-Dimensional Sparse Linear Contextual Bandits

We study the problem of dynamic batch learning in high-dimensional sparse linear contextual bandits, where a decision maker can only adapt decisions at a batch level. In particular, the decision maker, only observing rewards at the end of each batch, dynamically decides how many individuals to include in the next batch (at the current batchs end) and what personalized action-selection scheme to adopt within the batch. Such batch constraints are ubiquitous in a variety of practical contexts, including personalized product offerings in marketing and medical treatment selection in clinical trials. We characterize the fundamental learning limit in this problem via a novel lower bound analysis and provide a simple, exploration-free algorithm that uses the LASSO estimator, which achieves the minimax optimal performance characterized by the lower bound (up to log factors). To our best knowledge, our work provides the first inroad into a rigorous understanding of dynamic batch learning with high-dimensional covariates. We also demonstrate the efficacy of our algorithm on both synthetic data and the Warfarin medical dosing data. The empirical results show that with three batches (hence only two opportunities to adapt), our algorithm already performs comparably (in terms of statistical performance) to the state-of-the-art fully online high-dimensional linear contextual bandits algorithm. As an added bonus, since our algorithm operates in batches, it is orders of magnitudes faster than fully online learning algorithms. As such, our algorithm provides a desirable candidate for practical data-driven personalized decision making problems, where limited adaptivity is often a hard constraint.