improve docs and errors re: model formulas

NEWS.md

@@ -1,7 +1,10 @@
 # parsnip (development version)
 
+* Improved errors and documentation related to special terms in formulas. See `?model_formula` to learn more. (#770, #1014)
+
 * Improved errors in cases where the outcome column is mis-specified. (#1003)
 
+
 # parsnip 1.1.1
 
 * Fixed bug where prediction on rank deficient `lm()` models produced `.pred_res` instead of `.pred`. (#985)

R/gen_additive_mod.R

@@ -92,5 +92,21 @@ translate.gen_additive_mod <- function(x, engine = x$engine, ...) {
 #' @export
 #' @keywords internal
 fit_xy.gen_additive_mod <- function(object, ...) {
-  rlang::abort("`fit()` must be used with GAM models (due to its use of formulas).")
+  trace <- rlang::trace_back()
+
+  if ("workflows" %in% trace$namespace) {
+    cli::cli_abort(
+      c("!" = "When working with generalized additive models, please supply the
+               model specification to {.fun workflows::add_model} along with a \\
+               {.arg formula} argument.",
+        "i" = "See {.help parsnip::model_formula} to learn more."),
+      call = NULL
+    )
+  }
+
+  cli::cli_abort(c(
+    "!" = "Please use {.fun fit} rather than {.fun fit_xy} to train \\
+           generalized additive models.",
+    "i" = "See {.help model_formula} to learn more."
+  ))
 }

R/model_formula.R

@@ -0,0 +1,98 @@
+#' Formulas with special terms in tidymodels
+#'
+#' @description
+#'
+#' In R, formulas provide a compact, symbolic notation to specify model terms.
+#' Many modeling functions in R make use of ["specials"][stats::terms.formula],
+#' or nonstandard notations used in formulas. Specials are defined and handled as
+#' a special case by a given modeling package. For example, the mgcv package,
+#' which provides support for
+#' [generalized additive models][parsnip::gen_additive_mod] in R, defines a
+#' function `s()` to be in-lined into formulas. It can be used like so:
+#'
+#' ``` r
+#' mgcv::gam(mpg ~ wt + s(disp, k = 5), data = mtcars)
+#' ```
+#'
+#' In this example, the `s()` special defines a smoothing term that the mgcv
+#' package knows to look for when preprocessing model input.
+#'
+#' The parsnip package can handle most specials without issue. The analogous
+#' code for specifying this generalized additive model
+#' [with the parsnip "mgcv" engine][parsnip::details_gen_additive_mod_mgcv]
+#' looks like:
+#'
+#' ``` r
+#' gen_additive_mod() %>%
+#'   set_mode("regression") %>%
+#'   set_engine("mgcv") %>%
+#'   fit(mpg ~ wt + s(disp, k = 5), data = mtcars)
+#' ```
+#'
+#' However, parsnip is often used in conjunction with the greater tidymodels
+#' package ecosystem, which defines its own pre-processing infrastructure and
+#' functionality via packages like hardhat and recipes. The specials defined
+#' in many modeling packages introduce conflicts with that infrastructure.
+#'
+#' To support specials while also maintaining consistent syntax elsewhere in
+#' the ecosystem, **tidymodels delineates between two types of formulas:
+#' preprocessing formulas and model formulas**. Preprocessing formulas determine
+#' the model terms, while model formulas determine the model structure.
+#'
+#' @section Example:
+#'
+#' To create the preprocessing formula from the model formula, just remove
+#' the specials, retaining references to model terms themselves. For example:
+#'
+#' ```
+#' model_formula <- mpg ~ wt + s(disp, k = 5)
+#' preproc_formula <- mpg ~ wt + disp
+#' ```
+#'
+#' \itemize{
+#'   \item **With parsnip,** use the model formula:
+#'
+#'    ``` r
+#'    model_spec <-
+#'      gen_additive_mod() %>%
+#'      set_mode("regression") %>%
+#'      set_engine("mgcv")
+#'
+#'    model_spec %>%
+#'      fit(model_formula, data = mtcars)
+#'    ```
+#'
+#'    \item **With recipes**, use the preprocessing formula only:
+#'
+#'    ``` r
+#'    library(recipes)
+#'
+#'    recipe(preproc_formula, mtcars)
+#'    ```
+#'
+#'    The recipes package supplies a large variety of preprocessing techniques
+#'    that may replace the need for specials altogether, in some cases.
+#'
+#'   \item **With workflows,** use the preprocessing formula everywhere, but
+#'   pass the model formula to the `formula` argument in `add_model()`:
+#'
+#'    ``` r
+#'    library(workflows)
+#'
+#'    wflow <-
+#'      workflow() %>%
+#'      add_formula(preproc_formula) %>%
+#'      add_model(model_spec, formula = model_formula)
+#'
+#'    fit(wflow, data = mtcars)
+#'    ```
+#'
+#'    The workflow will then pass the model formula to parsnip, using the
+#'    preprocessor formula elsewhere. We would still use the preprocessing
+#'    formula if we had added a recipe preprocessor using `add_recipe()`
+#'    instead a formula via `add_formula()`.
+#'
+#' }
+#'
+#' @name model_formula
+NULL

_pkgdown.yml

@@ -78,6 +78,7 @@ reference:
       - control_parsnip
       - glance.model_fit
       - model_fit
+      - model_formula
       - model_spec
       - multi_predict
       - parsnip_addin

tests/testthat/test_gen_additive_model.R

@@ -26,7 +26,7 @@ test_that('regression', {
       y = mtcars$mpg,
       control = ctrl
     ),
-    regexp = "must be used with GAM models"
+    regexp = "to train generalized additive"
   )
   mgcv_mod <- mgcv::gam(mpg ~ s(disp) + wt + gear, data = mtcars, select = TRUE)
   expect_equal(coef(mgcv_mod), coef(extract_fit_engine(f_res)))
@@ -70,7 +70,7 @@ test_that('classification', {
       y = two_class_dat$Class,
       control = ctrl
     ),
-    regexp = "must be used with GAM models"
+    regexp = "to train generalized additive"
   )
   mgcv_mod <-
     mgcv::gam(Class ~ s(A, k = 10) + B,