Learning Optimal Interventions

In many data-driven applications such as medicine, the primary goal is identifying interventions that produce a desired change in some associated outcome. Such data is traditionally analyzed using models which facilitate human understanding of the relationships between variables. Based on conclusions drawn from this analysis, decision-makers choose interventions they confidently believe will improve outcomes. In contrast, we develop statistical algorithms that directly infer the best outcome-improving intervention. For applications where harmful intervention is drastically worse than proposing no change, we propose a conservative definition of the optimal intervention (under inherent uncertainty due to limited data). Taking into account practical constraints on which changes can actually be enacted, we consider both interventions which may be personalized based on the features of an individual as well as global policies applied uniformly over an entire population. Current applications of interest include: (1) Identifying changes to a online article's title that are likely to maximize the number of times the article is read, (2) Editing the amino-acid of an antibody to enhance its ability to target particular antigens.