Implement `fct_encode_one_hot()`

This PR aim to start the conversation on creating a function that takes a factor and return the corresponding integer indicator matrix. I find that I want to do this many places in tidymodels without a good way to do it.

This proposed function

returns integers
doesn't rely on global options
returns sensible column names
Works for factors that are too big for model.matrix()
Doesn't work with NAs

Benchmarking

This new function outperforms model.matrix() for all sizes of factors that I have tested

library(hardhat)

old_fun <- function(x) {
  res <- stats::model.matrix(~y-1, data = data.frame(y = x))
  attr(res, "assign") <- NULL
  attr(res, "contrasts") <- NULL
  res
}

create_factor <- function(n, n_levels) {
  factor(sample(seq_len(n_levels), n, replace = TRUE), levels = seq_len(n_levels))
}

res <- bench::press(
  n = 10 ^ seq_len(5),
  n_levels = 10 ^ seq_len(3),
  {
    fff <- create_factor(n, n_levels)
    bench::mark(
      new = unname(get_indicators(fff)),
      old = unname(old_fun(fff))
    )
  }
)

ggplot2::autoplot(res)

model.matrix() can't do factors with many levels


get_indicators(create_factor(10, 100000)) |> str()
#>  int [1:10, 1:100000] 0 0 0 0 0 0 0 0 0 0 ...
#>  - attr(*, "dimnames")=List of 2
#>   ..$ : NULL
#>   ..$ : chr [1:100000] "1" "2" "3" "4" ...
old_fun(create_factor(10, 100000))
#> Error: vector memory exhausted (limit reached?)

model.matrix() also couldn't calculate a vector with `10^8 elements and 20 levels.

res <- bench::press(
  n = 10 ^ seq_len(7),
  n_levels = 20,
  {
    fff <- create_factor(n, n_levels)
    bench::mark(
      min_iterations = 10,
      new = unname(get_indicators(fff)),
      old = unname(old_fun(fff))
    )
  }
)

ggplot2::autoplot(res)

^{Created on 2022-10-19 with reprex v2.0.2}

tidymodels / hardhat

Implement `fct_encode_one_hot()` #215

Benchmarking