Quantile regression integration

[topepo]

Closes #1125

Enables quantile regression models to be resampled and tuned.

Notes:

• The prediction summarization code can use some optimization but works fairly well.
• Takes parsnip and yardstick remotes. (EDIT)

An example:

library(tidymodels)
mirai::daemons(10)

set.seed(1)
sim_dat <- sim_regression(2000)
sim_rs <- vfold_cv(sim_dat)

bst_spec <-
  boost_tree(min_n = tune(), trees = 100, learn_rate = tune()) |>
  set_mode("quantile regression", quantile_levels = (0:10) / 10)

set.seed(2)
bst_res <-
  bst_spec |>
  tune_grid(
    outcome ~ .,
    resamples = sim_rs,
    control = control_grid(save_pred = TRUE, save_workflow = TRUE),
    grid = 25
  )

show_best(bst_res, metric = "weighted_interval_score")
#> # A tibble: 5 × 8
#>   min_n learn_rate .metric                .estimator  mean     n std_err .config
#>   <int>      <dbl> <chr>                  <chr>      <dbl> <int>   <dbl> <chr>  
#> 1    33     0.249  weighted_interval_sco… standard    4.73    10   0.113 pre0_m…
#> 2    19     0.316  weighted_interval_sco… standard    4.91    10   0.114 pre0_m…
#> 3    25     0.154  weighted_interval_sco… standard    5.00    10   0.132 pre0_m…
#> 4    11     0.196  weighted_interval_sco… standard    5.16    10   0.129 pre0_m…
#> 5    40     0.0953 weighted_interval_sco… standard    5.36    10   0.136 pre0_m…
bst_best <- select_best(bst_res, metric = "weighted_interval_score")
bst_best
#> # A tibble: 1 × 3
#>   min_n learn_rate .config         
#>   <int>      <dbl> <chr>           
#> 1    33      0.249 pre0_mod21_post0
fit_best(bst_res, metric = "weighted_interval_score")
#> ══ Workflow [trained] ══════════════════════════════════════════════════════════
#> Preprocessor: Formula
#> Model: boost_tree()
#> 
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#> outcome ~ .
#> 
#> ── Model ───────────────────────────────────────────────────────────────────────
#> ##### xgb.Booster
#> call:
#>   xgboost::xgb.train(params = list(eta = 0.248790236723884, max_depth = 6, 
#>     gamma = 0, colsample_bytree = 1, colsample_bynode = 1, min_child_weight = 33L, 
#>     subsample = 1, objective = "reg:quantileerror", nthread = 1, 
#>     quantile_alpha = c(0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 
#>     0.8, 0.9, 1)), data = x$data, nrounds = 100, evals = x$watchlist, 
#>     verbose = 0)
#> # of features: 20 
#> # of rounds:  100 
#> callbacks:
#>    evaluation_log 
#> evaluation_log:
#>   iter training_quantile
#>  <num>             <num>
#>      1          4.456474
#>      2          4.068782
#>    ---               ---
#>     99          1.046341
#>    100          1.041209

bst_res |> 
  collect_metrics() |> 
  ggplot(aes(learn_rate, min_n, size = mean)) +
  geom_point() + 
  scale_x_log10() +
  coord_fixed(1/15)

int_pctl(bst_res, parameters = bst_best, alpha = 0.1, times = 1000)
#> # A tibble: 1 × 8
#>   min_n learn_rate .metric            .estimator .lower .estimate .upper .config
#>   <int>      <dbl> <chr>              <chr>       <dbl>     <dbl>  <dbl> <chr>  
#> 1    33      0.249 weighted_interval… bootstrap    4.56      4.72   4.91 pre0_m…

^{Created on 2026-01-30 with reprex v2.1.1}

[topepo]

macOS error seems to be due to a CRAN issue with a package. I'll rerun periodically, but this is ready for review.

[EmilHvitfeldt]

looking very nice!

you also need to add a news bullet