The most important thing missing in documentation

[vzemlys]

So I have the recipe. What should I do next? How to pass the recipe to lm for example?

[treysp]

You process() it to return a tibble containing the transformed data, then pass that to your estimation function (e.g., lm()), as described in Basic Recipes. See also the process() reference.

[topepo]

So I have the recipe. What should I do next?

@treysp is correct but it is true that pushing the data that comes out of the recipe back into a formula is a bit underwhelming.

I'm in the process of making a recipe method for train and that should help a bit.

[vzemlys]

Cool. It is not clear that process is the last step. Especially since it is in documentation part called Preprocessing. Also learn is I think a bit of a misnomer, since correct me if I am wrong learning is the part where you apply the model, not the part where you preprocess data.

Sorry for nitpicking. Hope the feedback is useful.

[topepo]

Sorry for nitpicking. Hope the feedback is useful.

Absolutely!

It is not clear that process is the last step. Especially since it is in documentation part called Preprocessing

We're low on verbs to use 😬 . apply made more sense but it isn't a generic method.

I would say that process is the last step prior to model fitting (hopefully). The "pre-" is really pre-model fitting.

learn is I think a bit of a misnomer, since correct me if I am wrong learning is the part where you apply the model, not the part where you preprocess data.

The idea is that preprocessing of the data can be part of the overall modeling process and some of the steps involve estimating parameters in was that are similar to a classical model. For example, step_other needs to estimate the frequencies of class levels and so on. train was already taken by caret and some other packages.

The terminology is more similar to the machine learning communities where they would need to estimate/train/learn a model such as an autoencoder prior to fitting a neural network.

[treysp]

Maybe using verbs that imply an ordering would clarify the process for people outside machine learning communities (like me).

Although one could take it too far, what about leveraging the recipe metaphor? Maybe something like:

• First: create recipe object with recipe()
• Second: add steps to your recipe with step_*()
• Third: calculate intermediate elements such as frequencies of class levels with prep() (as in prepping ingredients)
• Fourth step = execute the recipe with cook() or something similar

Not sure how hokey that sounds 😁

[topepo]

prep is a pretty good word. Originally, I had thought of using cook or serve but we decided that we were stretching the food analogy at bit too far.

[treysp]

What about make for the last step? It doesn't seem as contrived as cook, and I doubt there would be any confusion with GNU Make.

[hadley]

I kind of like bake(). It rhymes with make() and is a bit more recipe like.