Quantifying Uncertainty of Uplift
Abstract: Uplift modeling refers to the task of estimating the causal effect of a treatment on an individual, also known as the conditional average treatment effect. Despite significant progress in uplift methods in recent years, the uncertainty of the estimates has been largely ignored in the literature. We explain why estimating uncertainty of the treatment effect is particularly important in many common use cases and we define epistemic uncertainty of the uplift estimates. Our main goal is to explain how uncertainty estimates can be incorporated into commonly used uplift model families with relative ease, and to demonstrate this we provide details for two practical methods that build on the standard approaches but add support for uncertainty quantification. We illustrate the methods on three real datasets and show how information about the uncertainty can be used in uplift modeling tasks, and additionally quantify the accuracy of the uncertainty estimates on simulated data.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: -New experiment comparing the estimated CIs with the true CATE on semi-simulated data.
-New experiment comparing the proposed models to three classic uplift models in terms of AUUC.
-Better positioning of the work within literature.
-Other minor corrections based on the reviews.
-Due to the requested changes, the paper now slightly exceeds 12 pages. However, it can easily be compressed back if needed.
Assigned Action Editor: ~Roman_Garnett1
Submission Number: 1163
Loading