Feature Request: Allow fit_xy for gen_additive_mod when not using the "mgcv" engine

[ben-e]

The current implementation of gen_additive_mod disables fit_xy due to the "mgcv" engine's use of formulas to specify smooths. I think this restriction is too mgcv specific, and may cause issues for other engines which specify smooths outside of the formula. Ideally, fit_xy could be disabled when engine = "mgcv", and enabled otherwise.

As an example, I am currently working on a package, wand, which allows users to specify smooths outside of a formula (example follows). I've currently got it running as its own model, nn_additive_mod, to work around the fit_xy restriction, but I think it would be a good candidate engine for gen_additive_mod.

# install.packages("pak")
# pak::pak("ben-e/wand")
library(wand)
library(modeldata)
library(parsnip)
library(recipes)
library(workflows)

data(bivariate)

# Using a workflow
wand_recipe <- recipe(Class ~ A + B, data = bivariate_train)
wand_recipe <- step_log(wand_recipe, all_numeric_predictors())

wand_spec <- nn_additive_mod(engine = "wand",
                             mode = "classification",
                             # Smooths are specified within a list, note that
                             # `B` here has already been log-transformed by
                             # the recipe.
                             smooth_specs = list(s_mlp(B)))

wand_wf <- workflow(preprocessor = wand_recipe, spec = wand_spec)

wand_fit <- fit(wand_wf, data = bivariate_train)

# Using a formula
wand_spec <- nn_additive_mod(engine = "wand",
                             mode = "regression",
                             # Rather than log-transforming with a recipe we
                             # can just do it in the smooth spec.
                             # If we want to specify the A ~ log(B) in the
                             # fit formula, then the smooth needs to be
                             # specified as s_mlp(`log(B)`).
                             smooth_specs = list(s_mlp(log(B))))

wand_fit <- fit(wand_spec,
                A ~ B,
                data = bivariate_train)

# This unfortunately does not translate well because nn_additive_mod is using
# the fit_xy interface which translates the formula to a dataframe prior to
# calling wand.
# As such, the fit formula can't accept something like A ~ s_mlp(log(B)),
# unlike gen_additive_mod which can accept a formula like `A ~ s(log(B))`
# In this case I'd just recommend users use wand directly:
# wand(A ~ s_mlp(log(B)))

I'd love to hear what ya'll think, thank you!

Ben

[topepo]

Sorry for the long wait.

We can remove that constraint on the overall class of models and just constrain the mgcv engine.

It will take me a few days to do this; currently in the weeds.

[github-actions]

This issue 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.