Outcome Prediction on Combinatorial Action Spaces

Abstract

Predicting outcomes of actions (treatments) to each target is a central task in decision support systems. For example, doctors decide whether to give a particular medication to individual patients aiming at better outcomes. In the outcome prediction problem, learning from biased observational data is a critical issue due to the lack of randomized controlled trial data in real situations. Recent efforts in causal inference and counterfactual machine learning attempt to estimate the impact of taking an action to a target. However, most of the existing work focuses only on a binary action space, and it is hard to handle a huge combinatorial action space, such as selecting an appropriate combination of medicines for a patient. To overcome this limitation, we formalize the outcome prediction as a domain adaptation task from a biased observational policy to a uniform random policy. The proposed representation balancing regularizer encourages the network to extract debiased representations from both the individual features and the combinations of actions. Our experimental results demonstrates the advantage of our method for combinatorial action spaces.