Shift machine learning backend to mlr3?

nt-williams / lmtp

:package: Non-parametric Causal Effects Based on Modified Treatment Policies :crystal_ball:

http://www.beyondtheate.com

GNU Affero General Public License v3.0

55 stars 16 forks source link

Shift machine learning backend to mlr3? #107

Open nt-williams opened 2 years ago

ck37 commented 10 months ago

I think this is a good idea - I’ve shifted my predictive work to mlr3, and piggybacking on it for tmle-related estimation would take advantage of their comparatively high software development capacity. I’m also planning to try out lmtp soon.

nt-williams commented 10 months ago

@ck37 Glad to hear you plan on trying out lmtp. A shift to mlr3 is tricky since the mlr3 API seems more suited to interactive use. I've been playing around with developing a convenience Super Learner package that uses mlr3 (https://github.com/nt-williams/mlr3superlearner) as a solution. Would be interested in your thoughts on this.

ck37 commented 10 months ago

The mlr3superlearner package looks awesome; I could see it really helping people, and the code efficiency is huge.

Just out of curiosity though, what is it that makes you feel mlr3 is best for interactive usage - the large amount of code needed to implement an ensemble?