Get correct coefs for ridge regression

[topepo]

closes #431

Makes a special parsnip parameter that allows users to set the penalty value independent of the full regularization path. This can help with pure ridge models where glmnet may not produce the correct values.

[juliasilge]

So are we not protecting against folks still doing the wrong thing in the penalty = 0 case? They will need to see this in the documentation to know what to do?

[topepo]

So are we not protecting against folks still doing the wrong thing in the penalty = 0 case? They will need to see this in the documentation to know what to do?

They will have to see the documentation to understand that.

We could fix the path values to a vector that would probably capture 90% of the cases (and document that). We can't effectively check the penalty value before setting a default; at parsnip predict-time, we don't know the range of penalties that will be used at predict-time.

The trade-off is:

• 👍 People won't silently get the wrong answers for ridge regression models.
• 👎 Model results might be slightly different than if we didn't set lambda; the default might cause poor results for penalties higher than our default.

Either is fine with me.

[juliasilge]

OK, I was fuzzy on our options before but I think I am clearer now.

I think we both agree that it is much more common for folks to want to use ridge regression than to end up needing very high penalty values. On the other hand, getting slightly different results than the underlying model is something that we know confuses people and that will happen basically for all cases of using glmnet ever if we set a path of lambdas.

This makes me think we shouldn't set a default. I am definitely a little worried that this is like a GOTCHA that people are not going to see; maybe in #456 we can think about how to highlight this kind of issue on, say, the glmnet landing page.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.