New step: step_dummy_manual()

[EmilHvitfeldt]

There are times where you have a limited number of levels in a factor, and you want to encode them in a specific way. Such as in the example below. Essentially it should perform a left_join() under the hood, but with special care taken with new and missing values.

library(recipes)

my_data <- tibble(
  choice = c("none", "A", "B", "both", "A", "none")
)

encoding <- tribble(
  ~value, ~A, ~B,
  "none",  0,  0,
  "A",     1,  0,
  "B",     0,  1,
  "both",  1,  1
)

recipe(~., data = my_data) %>%
  step_dummy_manual(choice, encoding = encoding) %>%
  prep() %>%
  bake(new_data = NULL)
#> # A tibble: 6 × 3
#>   choice_A    choice_B
#>      <dbl>       <dbl>
#> 1        0           0
#> 2        1           0
#> 3        0           1
#> 4        1           1
#> 5        1           0
#> 6        0           0

Another more realistic example would be if you had compas directions as an factor as we saw in https://www.kaggle.com/competitions/sliced-s01e04-knyna9

where you could do something like this:

encoding <- tribble(
  ~value, ~lon, ~lat,
  "N",       0,  1,
  "E",       1,  0,
  "S",       0, -1,
  "W",      -1,  0
)

weather example:

encoding <- tribble(
  ~value,          ~sun, ~cloud, ~rain,
  "sunny",            1,      0,     0,
  "partly clouded",   1,      1,     0,
  "clouded",          0,      1,     0,
  "raining",          0,      1,     1
)

[EmilHvitfeldt]

Also a way to handle out of table data, such as enriching. City names could be replaced/enriched with city characteristics