This repo contains a Python implementation of weighted doubly robust learning, due to Zhan, et al. (2024). It is a method of uplift modeling to recover causal effects of treatments in the presence of treatment confounding. For instance, a company may issue multiple coupons to individual customers during an ad campaign. This method, borrowing from Shapley values, is designed to disentangle each treatment's effect.
The two notebooks in the repo implement the algorithm, and demonstrate good performance on simulated data under different conditions.
The diagram below captures the essentials elements of the algorithm:
Zhan, B., Liu, C., Li, Y. and Wu, C., 2024. Weighted doubly robust learning: An uplift modeling technique for estimating mixed treatments' effect. Decision Support Systems, 176, p.114060. https://doi.org/10.1016/j.dss.2023.114060