API Move IterativeImputer outside from experimental - #34214
Conversation
|
ping @GaelVaroquaux if you have a little time to have a look at the documentation that I added and if it is inline with what you had in mind. |
There was a problem hiding this comment.
Thank you @glemaitre, I love cleanups! 🦾
It looks all good to me, only a nit comment and the suggestion to add recommendations for max_iter.
GaelVaroquaux
left a comment
There was a problem hiding this comment.
Great PR. I think that I'd like to add in the imputation docs and examples a few high-level summary messages on imputation for prediction.
Let me send you a PR to this PR :)
Co-authored-by: Stefanie Senger <[email protected]>
I rework the introduction with an approach that should be more towards what we would recommend nowadays given the understanding of the project. You can check the latest commit: 9ad4bc4 |
StefanieSenger
left a comment
There was a problem hiding this comment.
LGTM, only two little nits.
Thanks for your work.
Co-authored-by: Stefanie Senger <[email protected]>
Co-authored-by: Guillaume Lemaitre <[email protected]>
GaelVaroquaux
left a comment
There was a problem hiding this comment.
I tiny suggested change (on top of my own changes 😀 ), but IMHO this is ready to go.
Co-authored-by: Gael Varoquaux <[email protected]>
Co-authored-by: Stefanie Senger <[email protected]> Co-authored-by: Gael Varoquaux <[email protected]>
closes #14338
We never moved out of experimental
IterativeImputernotably because of the reason that we get convergence warning.However, the work in https://arxiv.org/abs/2407.19804 will tell us that we should not expect the iterative imputation to converge and that it can even be counter productive to spend too many imputation with a diminishing return.
Based on this research, we should make the
IterativeImputerout of experimental, stop to warn when it does not converge and instead make it such that we run it with a certain budget and document this positioning by recalling the literature.