search

uber / causalml

Uplift modeling and causal inference with machine learning algorithms
Other
5.02k stars 773 forks source link

Transformed Outcome methods #246

Closed shaddyab closed 3 years ago

shaddyab commented 3 years ago

Are there any plans to include Transformed Outcome algorithms in this package?

ppstacy commented 3 years ago

Hi @shaddyab thanks for using CausalML and bring this topic up. I don't think we have any plans right now. But we are open to anything for Causal Inference related algorithms. Do you already have something in mind or do you have any related papers or links to share with us? Thanks!

shaddyab commented 3 years ago

I know that the Transformed Outcome is already implemented in the pylift python package.I was wondering if it will also be included here with some adjusted modifications.

I also found this paper that presents a comparison between different algorithms with emphasize on revenue uplift Response transformation and profit decomposition for revenue uplift modeling

ppstacy commented 3 years ago

Thanks @shaddyab for sharing those!

Maybe I am wrong, but currently the predict() function in CausalML returns the predictions of treatment effects which are continuous values for both classification (pylift) and regression problem (revenue uplift). So I don't think we need to transform the outcome.

@jeongyoonlee @huigangchen @t-tte how do you guys think about it? Any outcome transfer method we should consider adding into the package?

shaddyab commented 3 years ago

@ppstacy , I agree that the current CausalML package already estimates the ATE; however, I thought it would be a useful additional to this package, especially that the transformed outcome is an unbiased estimate of the treatment effect. Hence, it can be used as a target variable for model training (i.e., A single meta learner is needed).

t-tte commented 3 years ago

As far as I understand the distinction between transformed outcome methods and others, our uplift tree methods fall under the category. But let me know if you have some other definition in mind.

More generally, if there are specific algorithms out there that have promising results in experiments, we're of course always interested in learning about them.