Research

Work in progress

Ethical Bandits (with Gui Liberali (RSM) and Yang Li (CKGSB))

Abstract: Upcoming legislation, such as the European GDPR, prohibit the use of sensitive consumer variables with the aim of reducing discrimination (fairness through unawareness). We find that such one-size-fits-all policies actually induce more discrimination and especially hurt minority groups legislators aim to protect. We propose a solution that ensures non-discrimination and adheres to ethical principles with respect to the distribution of consumer utility across consumer groups. 

MABs in Medical Trials: Heterogeneity and Delayed Outcome Observation (with Gui Liberali (RSM) and John Hauser (MIT))

Abstract: In this project we develop a multi-armed bandit approach for medical trials. We consider two specific problems in the field: 1) Personalized treatments, 2) handling delayed outcomes. Our method leverages biomarkers to solve both problems. First, we leverage stratification biomarkers, relating to the medical condition of the patient before treatment, for personalized treatment assignment. Second, we leverage reinforcement biomarkers, relating to the recovery of the patient, to update priors before the observation of the end-outcome. We analyze our methods on a random control trial on acute myocardial infarction (GUSTO-I).

Balancing Bandits: Data Aggregation in Adaptive Experiments  (with Gui Liberali (RSM))

Abstract: While highly personalized recommendations improve customer fit, they often suffer from data sparsity, leading to slow and noisy learning in multi-armed bandit (MAB) systems. Conversely, aggregating data across users accelerates learning but ignores individual preferences. This project investigates this personalization-aggregation trade-off. We propose a segmentation-based MAB framework that dynamically balances segment size (data volume) against within-segment heterogeneity (relevance) to optimize allocation of marketing interventions.