sample-props alternate implementation #20350
Conversation
There was a problem hiding this comment.
What do we generally want to do with estimators which are not meta-estimators per say, but they do have an estimator constructor arg, and forward {method}_params to fit or score. Should they do any kind of validation, or should they just pass everything along? Right now RFE does validation, but transformed regressor doesn't. I'm not really sure which way to go there.
| .fit_requests(sample_weight=True) | ||
| .score_requests(sample_weight=True) | ||
| ) | ||
| # The old behavior would be sample_weight=False for "score" | ||
| # Do we want to "fix" the issue, or keep the old behavior? |
There was a problem hiding this comment.
It seems to completely ignore sample_weight
|
The API side seems getting close to where we're happy with ii (@jnothman tell me if I'm wrong). I'm wondering how we can make this easier to move forward. Should we have a branch on the repo, where we start by merging the API/base stuff in a PR first, and then fixing estimators one by one, and then add some common tests? @glemaitre you showed interest in the last meeting, any help is welcome :) |
|
It'd also be nice if @thomasjpfan could check the API once again. |
|
With apologies for my general absence, that sounds like a good strategy for
breaking down the pull request.
I do consider TransformedTargetRegressor, RFE, etc to be meta estimators.
The meta estimator should be validating the metadata keys passed to it, but
not the values unless they are doing something like splitting the dataset.
Proper routing, aliasing and validation in these meta estimators ensures
capabilities like allowing them too to consume metadata unambiguously.
|
This is an alternate to #16079 which is an implementation of SLEP006 Routing sample-aligned meta-data.
The
metadata_requests.pyfile includes most of the implementation. There are some challenges left which we need to figure out how to tackle.You can see SLEP examples implemented in
test_props.pyastest_slep_case*functions, and the documentation undermetadata_routing.rstOpen Questions
getting_started.rstbe there? https://github.com/scikit-learn/scikit-learn/pull/20350/files#r659308253Closes #9566, Closes #15425, Closes #3524, Closes #15425,
Fixes #4497, Fixes #7136, Fixes #13432, Fixes #4632, Fixes #7646 (add test), Fixes #8127 (add test), Fixes #8158, Fixes #7308, Fixes #21134, Fixes #18159, Fixes #20349
Enables #6322,
TODO:
#12052, #8710, #15282, #2630, #8950, #11429, #15282, #18028, #19465, #20167
Note: since this started from where #16079 was, there are still leftovers from that PR, cleaning them up.