Giters

uber / causalml

Uplift modeling and causal inference with machine learning algorithms

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Add/update documentation

ras44 opened this issue · comments

commented

During Friday's office hours we discussed adding an introduction for new users about meta-learners, propensity modeling, etc. This issue was created to track and collaborate on those updates.

@jeongyoonlee, a couple questions:

  • just to confirm, should we focus on edits to the content in docs/*.rst?
  • any specific areas that you feel are in need of attention?

Thanks!

commented

Some notes on areas that could possibly use attention:

  • About CausalML (https://causalml.readthedocs.io/en/latest/about.html)

    • Is it Causal ML or CausalML?
    • use W for treatment variable for notation consistency
    • Add material on Neural network based algorithms
      • CEVAE
      • dragonnet

  • Methodology

    • Perhaps add a section on propensity score, since it is referenced frequently?

    • S-Learner

      • Stage 2
        • Here we fit the model with the entire dataset, including the treatment flag as a covariate
        • Possibly modify: 'Including the treatment propensity score...`
        • Perhaps "When the control and treatment groups have very different outcomes with respect to the covariates..."

    • T-Learner

      • In other words: we fit two models, each using the data from one of the two possible outcomes
      • define response under control, response under treatment \mu_0 and \mu_1

    • X-Learner

      • perhaps a high level intro mentioning the fact that X-Learners generate two estimates for CATE and then take a weighted average

    • R-Learner

      • add detail from Nie and Wager's paper

    • ... will continue to add

  • Installation

    • consider one single source of truth (currently this is replicated in README.md and the docs)
    • consider putting all install instructions in a separate markdown document that can include screenshots (helpful for the windows install)
      • using *.rst requires uploads of images into the repo, which bloats the repo size and is cumbersome
      • referencing, adding images, and editing a markdown within GitHub is relatively easy
      • possible to add docs/installation.md?

  • Examples

    • consider replacing all code chunks and images with links to jupyter notebooks or other sections

  • Interpretable Causal ML

    • consider replacing all code chunks and images with links to jupyter notebooks or other sections
commented

Thanks for taking the lead on this, @ras44.

To answer some of the questions:

  • docs/*.rst will be our main documentation published to readthedocs, while README.md will host a subset of instructions for installation and quick start examples.
  • Let's use CausalML to refer to the package.
  • I agree that we should use consistent variable naming conventions. e.g., W for treatment assignment

I will add more later.

commented

@jeongyoonlee reference to the User guide Flowchart we created in KDD 2021 tutorial, inspired by EconML's flowchart

external KDD 2021 - Introduction to CausalML (1)
external KDD 2021 - Introduction to CausalML

commented

hi @paullo0106 would it be possible to share a higher-res/sharper version of that second screenshot? I was hoping to add it to the documentation 🙏