RFC make shape of arrays consistent

[glemaitre]

Recently, we got some issues regarding the shape inconsistencies:

• of coef_ in linear model: TST add common tests for linear models for shape consistency of coef #20604
• of y_pred when providing a single target of shape (n_samples, 1): TST common test for predictions shape consistency with single target #20603

While it is pretty easy to make those arrays consistent across estimators, one could argue that the current user code is functioning.

This RFC is dedicated to know if we should solve this issue (see the failure in the different PRs) and if yes, shall we use a deprecation cycle to not break the code of our users from today to tomorrow?

[glemaitre]

ping @scikit-learn/core-devs

[NicolasHug]

Regarding y_pred, I'm in favour of always retuning y_pred.shape == (n_samples,) instead of (n_samples, 1). Quoting my #20603 (comment):

One important thing to note is that if we start adopting the output_shape == input_shape convention, then we can't use the y_pred.ndim == 1 iff is_single_target assumption anymore. We currently rely on that assumption quite a bit I think, sometimes implicitly as is the case for the original failure in #19352.

I haven't looked at the details yet, but I would probably follow the same logic for coef_

[glemaitre]

Regarding y_pred, I'm in favour of always retuning y_pred.shape == (n_samples,) instead of (n_samples, 1). Quoting my #20603 (comment):

Just to be complete, with a deprecation warning or with a breaking change @NicolasHug

[NicolasHug]

A deprecation should be possible and users can silence the warning by just passing a 1-D array.