TST add test to check that all ridge solver give the same results

[jeromedockes]

There lacks a test checking that all the ridge solvers (including the gcv) give similar results for a few different datasets (varying shapes, float32/64, and sparsity)

this test:

scikit-learn/sklearn/linear_model/tests/test_ridge.py

Line 290 in 0a2dc72

def test_ridge_individual_penalties():

compares solvers but for only one dense dataset with null coefficients and no intercept, and is meant to test individual penalties on outputs, so RidgeCV cannot be included since it does not support individual penalties.

after IRL with @ogrisel

[agramfort]

@jeromedockes do you time to also look at #13246

there is also there a problem with Ridge that is still different between dense and sparse

[jeromedockes]

@jeromedockes do you time to also look at #13246

there is also there a problem with Ridge that is still different between dense and sparse

yes, as far as I can tell the Ridge class (or other estimators that rely on it such as RANSAC with a Ridge as base estimator) is the only remaining problem for #13246 ?

It seems that at the moment, only the sparse_cg solver in Ridge can correctly fit an intercept when the design matrix is sparse, but it doesn't get selected when the solver is auto. SAG gets selected, and the documentation says it can fit an intercept with sparse X, but in most cases when X is sparse (hence not centered in preprocessing) sag requires many more iterations than the defaut max to find the right intercept, and the tolerance needs to be lowered as well.

the test that currently should check that Ridge fits an intercept uses a dataset where X has zero mean, which is why this problem is not caught by the tests on master.

I am planning to work on this and open a PR soon.

[NicolasHug]

Tests run pretty fast (about 1sec) so LGTM

[agramfort]

300 seeds seems a bit too much.

[jeromedockes]

my apologies about that, I had wanted to check with many seeds locally and
forgot to reduce the number of seeds afterwards. I set it to 3 now

[agramfort]

besides maybe a too ambitious/slow test LGTM. +1 for MRG

[agramfort]

Can I merge this one?

[rth]

Add copy=False

[rth]

Thanks @jeromedockes !

[jeromedockes]

Thanks @jeromedockes !

thanks for your help!