Can't we use Ridge(solver="cholesky", alpha=0) underneath in LinearRegression, to avoid many similar but not identical implementations? There may have been an issue with alpha=0 exactly though last time I checked.

I would be in favor of factoring the common code in a private helper function instead of having public estimators call into oneanother.

To summarize, this would involve the following tasks?

1. changing the scipy.linalg.lstsq call in LinearRegression to include the check_finite=False param
2. benchmarking lapack_driver choices for scipy.linalg.lstsq (should there also be an optional user param to choose lapack_driver?)
3. factoring out some Ridge code such that something like Ridge(solver="cholesky", alpha=0) can be used in LinearRegression without explicitly calling another estimator
4. determining when to use scipy.linalg.lstsq vs. cholesky—would this involve running experiments to find a good heuristic for tall/dense X?

hi @rithvikrao yes, I think that sounds right; apart from 1 maybe? Where did you get that from? We already have a global option to control that and in general we want to check for finiteness. Only if there's redundant checks would we want to remove them.

It's not entirely clear how easy 4 is, maybe starting with 3 would be good?

Hi @amueller, sounds good, I'll work on 3 first. 1 came from the first comment in this issue - I believe that fit in LinearRegression calls check_array, which is defined in utils/validation.py. This function by default raises an error on np.inf, np.nan, pd.NA in an array passed in. So I think the finiteness check is redundant, but I may be wrong there.

@rithvikrao sorry didn't see that from @ogrisel's original comment, you're right we can remove that check, that might even be the easiest first step.

This is related to #22855.