Weighted doubly robust learning: An uplift modeling technique for estimating mixed treatments' effect

Estimating the effect of mixed treatments is a crucial problem in causal inference. While previous studies have focused on econometric analysis, few have positioned the mixed treatment problem within the realm of causal machine learning, particularly in uplift modeling. This study proposes a novel uplift modeling technique called weighted doubly robust learning, which uses Shapley-value treatment attribution and doubly robust estimation to control for confounding among different treatments and estimate the pure effect for each treatment. Experiments are conducted on both synthetic dataset and industrial dataset. The results show that our method outperforms most of the current uplift modeling approaches in responsive customer targeting and effective treatment attribution, achieving an area under uplift curve (AUUC) of 0.590 and a Qini-coefficient of 0.080. Our method not only contributes to advancing current causal machine learning methods, but also provides valuable insights for companies in business decision making.