Personalization has long been central in machine learning, with successful applications to recommendation systems in online settings. A question of growing urgency is how to translate this success to emergent contexts where the data available has an inherent observational or behavioral nature, such as in personalized medicine. I will present a new machine learning toolset for building personalization models based on purely observational data, such as hospitals' electronic medical records (EMR), where the isolated effect of a treatment may be hidden by confounding factors. This is important because, unlike electronic and online settings, in medicine and other settings, experimentation can be prohibitively small-scale, costly, dangerous, and unethical in comparison to passive data collection, which can be massive. The toolset, which includes learning algorithms and validation schemes, is based on a new reformulation of the personalization problem as a single learning task and I will demonstrate empirically that it provides significant advantages over the standard approaches in specific personalized medicine and policymaking contexts. I will then present a particular application to personalized diabetes management, where we use EMRs from Boston Medical Center to devise a pharmacological treatment algorithm for their diabetes patients that provides personalized care based on patient characteristics, disease progression, and treatment history and achieves a medically and statistically significant 4.8 mmol/mol reduction in HbA1C (which measures blood glucose) relative to standard of care in the instances where it differs. Finally, I will present new results on dynamic assortment personalization in the face of a highly heterogenous population and very many items, where I show how to use low rank models and convex optimization to make low-regret learning from purely behavioral choice data practically feasible.

This talk includes joint work with D Bertsimas, M Udell, A Weinstein, and Y Zhou.