add step_lencode()

[EmilHvitfeldt]

Will close #232

library(embed)

n_obs <- 100000
n_lvls <- 100

data <- tibble(
  outcome = rnorm(n_obs),
  x = factor(sample(seq_len(n_lvls), n_obs, TRUE))
)

tictoc::tic("old")
res <- recipe(outcome ~ x, data = data) |>
  step_lencode_glm(x, outcome = vars(outcome)) |>
  prep()
tictoc::toc()
#> old: 1.157 sec elapsed

tictoc::tic("new")
res <- recipe(outcome ~ x, data = data) |>
  step_lencode(x, outcome = vars(outcome)) |>
  prep()
tictoc::toc()
#> new: 0.013 sec elapsed

We have yet to add regularization, but I'll add it as an argument in a future PR to keep things clean

[EmilHvitfeldt]

the big benefit for step_lencode() is that the performance is almost constant with respect to number of labels in the predictor.

library(embed)
n_obs <- 10000

res <- bench::press(
  n_lvls = c(5, 10, 20, 40, 80),
  {
    data <- tibble(
  outcome = rnorm(n_obs),
  x = factor(sample(seq_len(n_lvls), n_obs, TRUE))
)
    
    rec = recipe(outcome ~ x, data = data)
    bench::mark(
      check=FALSE,
      old = rec |> step_lencode_glm(x, outcome = vars(outcome)) |> prep(),
      new = rec |> step_lencode(x, outcome = vars(outcome)) |> prep()
    )
  }
)
#> Running with:
#>   n_lvls
#> 1      5
#> 2     10
#> 3     20
#> 4     40
#> 5     80
library(bench)
library(ggplot2)

res |>
  ggplot(aes(n_lvls, median, color = expression)) +
  geom_point() +
  scale_color_bench_expr() +
  scale_x_log10()

res |>
  ggplot(aes(n_lvls, mem_alloc, color = expression)) +
    geom_point() +
  scale_color_bench_expr() +
    scale_x_log10()

^{Created on 2025-08-27 with reprex v2.1.1}

[EmilHvitfeldt]

and it performs better with respect to the number of rows

library(embed)

n_lvls <- 10

res <- bench::press(
  n_obs = c(100, 1000, 10000, 100000),
  {
    data <- tibble(
  outcome = rnorm(n_obs),
  x = factor(sample(seq_len(n_lvls), n_obs, TRUE))
)
    
    rec = recipe(outcome ~ x, data = data)
    bench::mark(
      check=FALSE,
      old = rec |> step_lencode_glm(x, outcome = vars(outcome)) |> prep(),
      new = rec |> step_lencode(x, outcome = vars(outcome)) |> prep()
    )
  }
)
#> Running with:
#>    n_obs
#> 1    100
#> 2   1000
#> 3  10000
#> 4 100000
#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
library(bench)
library(ggplot2)

res |>
  ggplot(aes(n_obs, median, color = expression)) +
  geom_point() +
  scale_color_bench_expr() +
  scale_x_log10()

res |>
  ggplot(aes(n_obs, mem_alloc, color = expression)) +
    geom_point() +
  scale_color_bench_expr() +
    scale_x_log10()

^{Created on 2025-08-27 with reprex v2.1.1}