Input data for workflows

[topepo]

Related to #99 and #118

A workflow combines a pre-processer (e.g. an R formula) as well as a model. The pre-processor can also include a recipe.

Because of this, vip might want to be able to differentiate between the original columns and the features. The workflow takes the original columns and the model takes the features.

Right now, if we want to get the importances of the original columns, I didn't think that we can. For example, if a recipe converts some of the original columns to principal components, the user probably wants the importances of the variables and not the PCA features.

I think that we could make a specific methods for workflows for vi_firm(), vi_ice(), vi_permute(), and vi_shap() (did I miss anything?). These methods could have an option for the user to pick between "original" or "derived" predictors. I don't know if that messes up any of the other components of the package.

This would mean that vip would take a workflows dependency (which isn't so bad I think).

If you are interested, let me know and I'll do a PR.

[bgreenwell]

I'm not up to speed on {workflows} yet, but if I understand it correctly, this has always been possible in vi_permute(), vi_firm() (which evolved from the older vi_pdp() and vi_ice()), and vi_shap() by supplying the "right" prediction wrapper. Since the documentation/examples here don't quite show how to accomplish variable importance in the original (un-processed) features, I think adding specific methods would be great. Would be delighted if you made a PR :). Would {workflows} necessarily be a dependency here though, or just included in the Suggests field?

[bgreenwell]

@topepo, does your recent PR address this particular issue?