Causal Effect Evaluation for Decision-Making in Large Treatment Space

Abstract

Predicting which intervention (actions) will lead to better outcomes is a central task for decision support systems. To build predictive models in real-world settings, we have to learn from observational data with sampling bias due to the lack of randomized controlled trial (RCT) data. For this problem, recent efforts in causal inference and counterfactual machine learning have focused on estimating the potential outcomes and their differences on a binary action space such as whether to prescribe medication. However, when it comes to large action spaces (e.g., selecting an appropriate combination of medication for an individual patient), the regression accuracy of the potential outcome is no longer sufficient to obtain sufficient decision-making performance practically as shown. The proposed loss function improves the decision-making performance based on the learned model by minimizing the classification error of whether the action is relatively better than the past average decision maker's (doctor's) action for an individual situation (patient). Experiments on a semi-synthetic dataset demonstrate the superiority of the proposed method for large action spaces.