Multi-treatment casual analysis using improved meta learner and uplift tree

Abstract

Causality is an appealing but challenging domain for researchers in generations. Recently, researchers have shifted their focus to combining traditional causal inference methods and machine learning models to get both advantages. Meta learner is an algorithm for causal inference, including T-learner, S-learner, and X-learner. Another popular way in causal inference is based on decision tree learning, one of the predictive modeling approaches. Many existing works focus on estimating the causal effect of binary treatment. However, there are also many cases in the real world when the treatment has more than two values. These methods cannot be used directly in multivalued treatment cases. According to the mathematization of causality, we improved the binary meta-learner process to be applicable in multi-treatment situations. At the same time, we also preliminarily explored the technique of uplifting trees. Finally, we applied the two methods to analyze parents' and children's learning situations in hundreds of families to test the effect of improvement.