Learning to Personalize from Observational Data

Nathan Kallus

Published 2016-08-31Version 1

We study the problem of learning to choose from m discrete treatment options (e.g., medical drugs) the one with best causal effect for a particular instance (e.g., patient) characterized by an observation of covariates. The training data consists of observations of covariates, treatment, and the outcome of the treatment. We recast the problem of learning to personalize from these observational data as a single learning task, which we use to develop four specific machine learning methods to directly address the personalization problem, two with a unique interpretability property. We also show how to validate personalization models on observational data, proposing the new coefficient of personalization as a unitless measure of effectiveness. We demonstrate the power of the new methods in two specific personalized medicine and policymaking applications and show they provide a significant advantage over standard approaches.