Q: Data first in the pipeline?

[njtierney]

Hi there!

Thanks so much for your work so far on this, I'm really looking forward to seeing this package develop!

I was wondering if you have any thoughts on placing data as the first argument in a recipe?

This would then mean that the README code could then be:

sonar_rec <- Sonar %>%
  recipe(Class ~ .) %>%
  step_center(all_predictors()) %>%
  step_scale(all_predictors())

This idea has been made possible with the intubate package, but perhaps this could be something provided in recipes?

Thank you again for all your work!

[topepo]

The formula comes first so that that it is an S3 formula method for the recipe class. Let me see if there is a fancy way to do this without resorting to S4 methods.

[njtierney]

OK, sounds good! :)

[topepo]

Some thinking quasi out loud about this...

If we were to do this, we would get rid of the formula method for the recipe class and use the default method to direct arguments to places that they should go.

The current class system has:

recipe <- function(x, ...) UseMethod("recipe")

## `...`` are not currently hooked up to anything

recipe.default <- function(x, vars = colnames(x), roles = NULL, ...) {
  # code
}

recipe.formula <- function(formula, data, ...) {
  # code that eventually calls `recipe.default`
}

If the function is called in the usual way with a formula:

recipe(y ~ x, data = biomass_tr)

but there is no formula method, this ends up being

# case 1
recipe.default(x = y ~ x, data = biomass_tr)

Note that data is now part of ... and would need to be intercepted and reassigned.

If called from a pipe:

data %>% recipe(y ~ x)

the default method would result in

# case 2
recipe.default(x = biomass_tr, vars = y ~ x)

In both cases, we would need to pass to some other non-method function (say recipe_form) that executes the workflow that the current recipe.formula has and then passes arguments to recipe.default.

It seems pretty ugly since, in case 1, x is nothing like what we would claim in the man page. I could change it to something like object but that is pretty kludgy since I don't think that anyone would expect something called object to be the data (in R, anyway).

Is there a more elegant way to get to the same place?

[hadley]

Why not something like this:

recipe <- function(data, formula = NULL, ...) UseMethod("recipe")

recipe.data.frame <- function(x, formula = NULL, ..., vars = colnames(x), roles = NULL) {
  
  recipe <- ...

  if (!is.null(formula))
    recipe <- step_formula(recipe, formula)
  
}

Then recipe(data, formula) is just a convenient short hand for data %>% recipe() %>% step_formula(formula)

[topepo]

That should solve it but please test with your problem