Repo Reviews

Review

• Author: @ras44
• Repo: uber/causalml
• Title: Best to build from source, some dep issues, econml alternative

Benefits

• CausalML provides a suite of state-of-the-art machine learning tools for estimating treatment effects
• Applications include campaign optimization, personalization
• README contains details on available algorithms
• Quickstart contains examples and includes interpretability, vizualization features
• We have used causalml successfully for production causal inference

Challenges

• Active development has slowed as of 2022 as multiple original contributors have moved on

• Repo is being incubated, but it’s unclear if that’s for quality assurance or as a sign of end of support/development

• The repo contains a variety of environment files in envs/environment[-tf]-py[36,37,38] and multiple users experience issues creating conda environments with them

  • I’ve had the most success just building from source, for example:

conda create -n py38 python=3.8
conda activate py38
cd causalml
python -m pip install .[test]
pytest -vs tests/ --cov causalml/

• However, if building from source, there are important constraints to consider that can be found in the pyproject.toml file:

  • Here we see that we are locked into scikit-learn<1.0.2 and numpy<1.23, which can lead to conflicts with more recently updated packages.

  • Releasing these constraints would require significant lift

• Note that you can at times create the conda environments, but not all tests will run successfully.
  • This is what leads to most of the issues mentioning errors in the conda env installs
• An alternative repo that seems to have much more activity lately is https://github.com/py-why/EconML

```