mglisse
commented
2 weeks ago According to their doc, the goal of this feature is to be able to pass extra arguments (not just X and y) to fit or transform, when they are not called directly but through a pipeline, grid-search, or other similar meta-objects.
The classes listed in this issue seem to be precisely those for which fit takes extra arguments (sample_weight in Atol, filter(s), color(s), assignments in the cover complex classes), i.e. the ones that could benefit.
The developer doc gives more information about how this works and what may be needed in our classes to enable it.
The whole feature is marked as experimental and likely to break at each release, so we probably shouldn't try to support it yet. I would be in favor of option 3 above, for now.
With recent (1.3+) versions of scikit-learn, building the documentation prints
for Atol, MapperComplex, GraphInducedComplex and NerveComplex. New functions also get added to the documentation:
get_metadata_routing,set_fit_request,set_transform_request. I don't know what criteria cause these to appear for some classes and not others.Explanation is available at https://github.com/scikit-learn/scikit-learn/pull/26747
Some options for what we can do
We should probably check first what this new metadata_routing thing is about and whether it is relevant to us (if it is, we may want code changes, not just the doc).