step_other creates a specification of a recipe
step that will potentially pool infrequently occurring values
into an "other" category.
step_other(
recipe,
...,
role = NA,
trained = FALSE,
threshold = 0.05,
other = "other",
objects = NULL,
skip = FALSE,
id = rand_id("other")
)| recipe | A recipe object. The step will be added to the sequence of operations for this recipe. |
|---|---|
| ... | One or more selector functions to choose which
variables that will potentially be reduced. See
|
| role | Not used by this step since no new variables are created. |
| trained | A logical to indicate if the quantities for preprocessing have been estimated. |
| threshold | A numeric value between 0 and 1 or an integer greater or
equal to one. If it's less than one then factor levels whose rate of
occurrence in the training set are below |
| other | A single character value for the "other" category. |
| objects | A list of objects that contain the information
to pool infrequent levels that is determined by
|
| skip | A logical. Should the step be skipped when the
recipe is baked by |
| id | A character string that is unique to this step to identify it. |
An updated version of recipe with the new step
added to the sequence of existing steps (if any).
The overall proportion (or total counts) of the categories are
computed. The "other" category is used in place of any categorical levels
whose individual proportion (or frequency) in the training set is less than
threshold.
If no pooling is done the data are unmodified (although character data may
be changed to factors based on the value of strings_as_factors in
prep.recipe()). Otherwise, a factor is always returned with
different factor levels.
If threshold is less than the largest category proportion, all levels
except for the most frequent are collapsed to the other level.
If the retained categories include the value of other, an error is
thrown. If other is in the list of discarded levels, no error
occurs.
If no pooling is done, novel factor levels are converted to missing. If pooling is needed, they will be placed into the other category.
When data to be processed contains novel levels (i.e., not contained in the training set), the other category is assigned.
When you tidy() this step, a tibble with columns terms (the
columns that will be affected) and retained (the factor
levels that were not pulled into "other") is returned.
step_factor2string(), step_string2factor(),
dummy_names(), step_regex(), step_count(),
step_ordinalscore(), step_unorder(), step_novel()
library(modeldata)
data(okc)
set.seed(19)
in_train <- sample(1:nrow(okc), size = 30000)
okc_tr <- okc[ in_train,]
okc_te <- okc[-in_train,]
rec <- recipe(~ diet + location, data = okc_tr)
rec <- rec %>%
step_other(diet, location, threshold = .1, other = "other values")
rec <- prep(rec, training = okc_tr)
collapsed <- bake(rec, okc_te)
table(okc_te$diet, collapsed$diet, useNA = "always")
#>
#> anything mostly anything strictly anything other values
#> anything 3128 0 0 0
#> halal 0 0 0 4
#> kosher 0 0 0 5
#> mostly anything 0 8226 0 0
#> mostly halal 0 0 0 23
#> mostly kosher 0 0 0 36
#> mostly other 0 0 0 509
#> mostly vegan 0 0 0 164
#> mostly vegetarian 0 0 0 1715
#> other 0 0 0 172
#> strictly anything 0 0 2562 0
#> strictly halal 0 0 0 7
#> strictly kosher 0 0 0 7
#> strictly other 0 0 0 209
#> strictly vegan 0 0 0 104
#> strictly vegetarian 0 0 0 442
#> vegan 0 0 0 66
#> vegetarian 0 0 0 331
#> <NA> 0 0 0 0
#>
#> <NA>
#> anything 0
#> halal 0
#> kosher 0
#> mostly anything 0
#> mostly halal 0
#> mostly kosher 0
#> mostly other 0
#> mostly vegan 0
#> mostly vegetarian 0
#> other 0
#> strictly anything 0
#> strictly halal 0
#> strictly kosher 0
#> strictly other 0
#> strictly vegan 0
#> strictly vegetarian 0
#> vegan 0
#> vegetarian 0
#> <NA> 12145
tidy(rec, number = 1)
#> # A tibble: 5 x 3
#> terms retained id
#> <chr> <chr> <chr>
#> 1 diet anything other_ez6XT
#> 2 diet mostly anything other_ez6XT
#> 3 diet strictly anything other_ez6XT
#> 4 location oakland other_ez6XT
#> 5 location san francisco other_ez6XT
# novel levels are also "othered"
tahiti <- okc[1,]
tahiti$location <- "a magical place"
bake(rec, tahiti)
#> # A tibble: 1 x 2
#> diet location
#> <fct> <fct>
#> 1 strictly anything other values
# threshold as a frequency
rec <- recipe(~ diet + location, data = okc_tr)
rec <- rec %>%
step_other(diet, location, threshold = 2000, other = "other values")
rec <- prep(rec, training = okc_tr)
tidy(rec, number = 1)
#> # A tibble: 6 x 3
#> terms retained id
#> <chr> <chr> <chr>
#> 1 diet anything other_p2QWY
#> 2 diet mostly anything other_p2QWY
#> 3 diet strictly anything other_p2QWY
#> 4 location berkeley other_p2QWY
#> 5 location oakland other_p2QWY
#> 6 location san francisco other_p2QWY
# compare it to
# okc_tr %>% count(diet, sort = TRUE) %>% top_n(4)
# okc_tr %>% count(location, sort = TRUE) %>% top_n(3)