Remove warnings from skore.train_test_split

[auguste-probabl]

What would you like to say?

Many examples and tests end up using sklearn's train_test_split because we know that skore's warnings are not relevant for the task at hand. We know that the warnings experience is sub-optimal in most cases.

On the other hand, skore.train_test_split innovates through the as_dict parameter, which would make train-test-splitting easier in many cases; so removing skore.train_test_split would be throwing out nice functionality.

I would like to be able to use skore.train_test_split without having to worry about warnings. To solve this, I propose to add a warn flag, False by default, which controls whether data checks are run or not.

[MarieSacksick]

If inside our own examples we don't use our own function because it throws un-relevant warnings, doesn't it mean that we should change the warnings instead?
(I learned a new term this morning, the XY problem, and it just looks exactly that).

[auguste-probabl]

By "examples" I mostly meant doctests, not the full-blown examples. The issue is that in these doctests, the point is not to show the warnings, it's to showcase some other skore capability; so these warnings are parasitic, no matter what they might say. A fine way to solve this issue could also be to add warnings.catch_warnings() to hide the warnings.

[MarieSacksick]

I prefer your second solution. Enabling to mute warnings in the function itself would mean that there are actual real-life use-cases, that are not too rare, where it's actually what should be done.

[auguste-probabl]

Thinking a bit more about it, I think the warnings have no place in train_test_split. I would like them moved out to a new bit of functionality, like what was discussed in #1509, or removed until we can think of a better UX (e.g. something that integrates in a TableReport).