Closed samcarlos closed 4 years ago
Thanks Sam, I think we will probably create a subsection (under the "learning causal effects" section) for uplift modeling later, as methodologically it is kind of different from effect estimation, especially in term of the evaluation metrics. If you can contribute that for the repo it would also be nice.
Thanks,
Ruocheng
Thanks @rguo12! My understanding is that "Uplift Models" is just another way of saying heterogenous treatment effects or Individual treatment effect estimations. This package uses similar methodologically to estimation of the treatment effect as the BART Paper by Hill in that section.
How would you say uplift models are distinct from HETE or ICE models? Thanks, Sam
You are right, they are essentially the same problem, with CATE accurately estimated, the problem is solved. Just like in offline policy evaluation, if you estimate counterfactual outcomes/CATE correctly, the problem is solved, although it is recommended to be solved by doubly robust estimators, that is to emphasize the importance of propensity score modeling in the offline policy evaluation problem.
Maybe I did not read many uplift modeling papers, but IMO uplift modeling often evaluates models with the uplift curves or Qini curves where you check the gain of assigning treatments to arbitrary subpopulations (percentiles). The x axis is interesting here as it sorts instances according to the causal effects. Looking at those curves, maybe the focus is to develop a strategy to assign treatments when we only can assign treatments to a limited amount of people.
Please correct me if I am wrong, since I have not read many uplift modeling papers.
Best regards,
Ruocheng
On Wed, Apr 22, 2020, 3:08 PM Samuel Weiss notifications@github.com wrote:
Thanks @rguo12 https://github.com/rguo12! My understanding is that "Uplift Models" is just another way of saying heterogenous treatment effects or Individual treatment effect estimations. This package uses similar methodologically to estimation of the treatment effect as the BART Paper by Hill in that section.
How would you say uplift models are distinct from HETE or ICE models? Thanks, Sam
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/rguo12/awesome-causality-algorithms/pull/5#issuecomment-618066016, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABUVWHOOE25C55VBFIAAANDRN5TFPANCNFSM4MOMCE4A .
Hey Ruocheng,
Yea I think you're right. My understanding that a bunch of different domains solved ITE problems separately and came up with different terminology. Kind of like how some people say 'feature' and others say 'independent variable'. The use cases might be slightly different but the tools and general theory is widely applicable between them.
Thanks, Sam
Thanks Sam, it is very nice to discuss with you. I will keep updating the repo when I have time. I really appreciate your contribution.
Best regards,
Ruocheng
On Wed, Apr 22, 2020, 3:48 PM Samuel Weiss notifications@github.com wrote:
Hey Ruocheng,
Yea I think you're right. My understanding that a bunch of different domains solved ITE problems separately and came up with different terminology. Kind of like how some people say 'feature' and others say 'independent variable'. The use cases might be slightly different but the tools and general theory is widely applicable between them.
Thanks, Sam
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/rguo12/awesome-causality-algorithms/pull/5#issuecomment-618079268, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABUVWHOC4C6X3WAVHWLJL43RN5X3VANCNFSM4MOMCE4A .
Nice talking with you Ruocheng. Thanks!
On Wed, Apr 22, 2020 at 5:55 PM Ruocheng Guo notifications@github.com wrote:
Thanks Sam, it is very nice to discuss with you. I will keep updating the repo when I have time. I really appreciate your contribution.
Best regards,
Ruocheng
On Wed, Apr 22, 2020, 3:48 PM Samuel Weiss notifications@github.com wrote:
Hey Ruocheng,
Yea I think you're right. My understanding that a bunch of different domains solved ITE problems separately and came up with different terminology. Kind of like how some people say 'feature' and others say 'independent variable'. The use cases might be slightly different but the tools and general theory is widely applicable between them.
Thanks, Sam
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub < https://github.com/rguo12/awesome-causality-algorithms/pull/5#issuecomment-618079268 , or unsubscribe < https://github.com/notifications/unsubscribe-auth/ABUVWHOC4C6X3WAVHWLJL43RN5X3VANCNFSM4MOMCE4A
.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/rguo12/awesome-causality-algorithms/pull/5#issuecomment-618114806, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAPGINYWO7X4XLM35UAJRK3RN6GZHANCNFSM4MOMCE4A .
I would like to include my MR-Uplift package. This package a models an uplift model for multiple response variables and evaluates tradeoffs between them. Thanks!