I added H2O's Uplift Random Forest estimator as a new method. This tree-based algorithm was proposed by Rzepakowski & Jaroszewicz (2012) [1] and comes with three different uplift-specific splitting criteria: Kullback-Leibler, Euclidean Distance, and Chi-squared divergence.

I used H2O's H2OUpliftRandomForestEstimator and

1. Implemented an UpliftRandomForestEstimator class
2. Added tests for the UpliftRandomForestEstimator
3. Added documentation for the UpliftRandomForestEstimator

[1] Rzepakowski, P., & Jaroszewicz, S. (2012). Decision trees for uplift modeling with single and multiple treatments. Knowledge and Information Systems, 32(2), 303-327.

I added stationary feature selection methods, that is:

• Divergence Filter (Filter method) [1]
• Net Information Value (Filter method) [2]
• Uplift Curve (Filter method) [3]
• Permutation with Uplift Random Forest (Wrapper method)

[1] Zhao, Z., Zhang, Y., Harinen, T., & Yung, M. (2022). Feature Selection Methods for Uplift Modeling and Heterogeneous Treatment Effect. In IFIP International Conference on Artificial Intelligence Applications and Innovations (pp. 217-230). Springer, Cham. [2] Larsen, K. (2015). Data Exploration with Weight of Evidence and Information Value in R. Retrieved Date from https://multithreaded. stitchfix. com/blog/2015/08/13/weight-of-evidence. [3] Hu, J. (2022). Customer feature selection from high-dimensional bank direct marketing data for uplift modeling. Journal of Marketing Analytics, 1-12.

I noticed Vim temporary .swp files are being tracked by git. I would add:

• .swp (Vim working buffer)
• *~ (common backup for Linux based editor)
• some other common editor artifact I'm missing

Removing the line that creates the "cvt_label" column in the "predict" method of the CVT Model as it is not used in this method (I guess it was copied by mistake from the "fit" method) and will eventually prevent the method from being used for the inference stage of an actual machine learning model.

First of all, thank you for developing this awesome package! PySpark/H2O-based algorithms are like gospels for real world large marketing datasets! Liked Uber's CausalML few years ago, but their implementation is in Pandas/sklearn so we were not able to use it on massive dataset. This package solved that issue perfectly!

Recently I was reading the tutorial on uplift model posted by your team for Web Conference 2021 (WWW’21)(https://booking.ai/uplift-modeling-f9759e3fb51e). At the end of the deck I read about the recent paper of multiple treatment with cost optimization by Uber(https://arxiv.org/abs/1908.05372), which extend the existing R-/X-learner to a multi-treatment setting, with/without control.

If I'm not mistaken, the upliftml package currently can only handle one treatment and on control scenerio for meta-learners. Want to know if your team has any development plan to extend meta-learner to multi-treatment with/without control group? Thanks!

The signatures of our evaluation functions are largely overlapping, in particular when it comes to bootstrapping-related parameters. When changing one of these parameters (e.g. adding another bootstrapping-related param), the current implementation requires one to change the signatures of all of the evaluation functions manually. To reduce this duplication work, we could lump the bootstrapping-related parameters together as bootstrap_kwargs, e.g. eval_function(arg1, arg2, **bootstrap_kwargs).

Add use_std_error argument to all procedures that compute a confidence interval through bootstrapping. Default to True because this method is more efficient and more familiar.

Add explanation and retouch literature reference in docstring

Solution to #1

Missing: tests do not cover both computation methods as they are written now. Writing tests for every metric would take a lot of duplication.

The current implementation estimates the confidence interval directly from bootstraps. Another option would be to use bootstrapped values to get the standard error and then derive the CI from normal (or student's T) critical values. The former has worse statistical properties than the latter (you need more iterations for it to converge), but the latter expects the distribution of the estimator to be symmetric.

We could implement the std error based solution, set it as default, but keep an option for the user to switch back to the direct CI estimation.

Thanks @ilirmaci for pointing it out!