Orateur
Description
In domains such as personalized medicine, historical data is used to learn what treatments to prescribe to maximize positive outcomes. Previous studies have proposed methods for creating prescriptive trees: human-interpretable diagrams that indicate what type of treatment an individual should get based on their measurements. However, a remaining problem is that the models perform worse over time when there is a shift in the data collection process or when data from a different source is used during the training and prediction phases. To solve this problem, we propose a method that considers data uncertainty by optimizing distributionally robust prescriptive trees. We formulate a linear-time algorithm to find the worst-case distribution shift within a given Wasserstein distance around the dataset and use it as a subroutine within the main problem. Our algorithm does not depend on any specific causal effect estimator and can, therefore, be applied in various contexts.
