recipes (version 0.1.16)

selections: Methods for Selecting Variables in Step Functions

Description

When selecting variables or model terms in step functions, dplyr-like tools are used. The selector functions can choose variables based on their name, current role, data type, or any combination of these. The selectors are passed as any other argument to the step. If the variables are explicitly stated in the step function, this might be similar to:

  recipe( ~ ., data = USArrests) %>%
    step_pca(Murder, Assault, UrbanPop, Rape, num_comp = 3)

The first four arguments indicate which variables should be used in the PCA while the last argument is a specific argument to step_pca().

Note that:

1. These arguments are not evaluated until the prep function for the step is executed.

2. The dplyr-like syntax allows for negative signs to exclude variables (e.g. -Murder) and the set of selectors will processed in order.

3. A leading exclusion in these arguments (e.g. -Murder) has the effect of adding all variables to the list except the excluded variable(s).

Select helpers from the tidyselect package can also be used: tidyselect::starts_with(), tidyselect::ends_with(), tidyselect::contains(), tidyselect::matches(), tidyselect::num_range(), tidyselect::everything(), tidyselect::one_of(), tidyselect::all_of(), and tidyselect::any_of()

For example:

  recipe(Species ~ ., data = iris) %>%
    step_center(starts_with("Sepal"), -contains("Width"))

would only select Sepal.Length

Columns of the design matrix that may not exist when the step is coded can also be selected. For example, when using step_pca(), the number of columns created by feature extraction may not be known when subsequent steps are defined. In this case, using matches("^PC") will select all of the columns whose names start with "PC" once those columns are created.

There are sets of recipes-specific functions that can be used to select variables based on their role or type: has_role() and has_type(). For convenience, there are also functions that are more specific. The functions all_numeric() and all_nominal() select based on type, with nominal variables including both character and factor; the functions all_predictors() and all_outcomes() select based on role. Any can be used in conjunction with the previous functions described for selecting variables using their names:

  data(biomass)
  recipe(HHV ~ ., data = biomass) %>%
    step_center(all_numeric(), -all_outcomes())

This results in all the numeric predictors: carbon, hydrogen, oxygen, nitrogen, and sulfur.

If a role for a variable has not been defined, it will never be selected using role-specific selectors.

Selectors can be used in step_interact() in similar ways but must be embedded in a model formula (as opposed to a sequence of selectors). For example, the interaction specification could be ~ starts_with("Species"):Sepal.Width. This can be useful if Species was converted to dummy variables previously using step_dummy(). The implementation of step_interact() is special, and is more restricted than the other step functions. Only the selector functions from recipes and tidyselect are allowed. User defined selector functions will not be recognized. Additionally, the tidyselect domain specific language is not recognized here, meaning that &, |, !, and - will not work.