support "censored regression"

[simonpcouch]

Will need to come after workflows/workflowsets support, but--

A good bit of conditional statements based on the mode will need to be adjusted, to start off!🏄

stacks/R/add_candidates.R

Lines 185 to 186 in 227efb6

	# note whether classification or regression
	.set_mode_ <- function(stack, candidates, name) {

stacks/R/blend_predictions.R

Line 158 in 7822d86

if (attr(data_stack, "mode") == "regression") {

stacks/R/print.R

Line 5 in 0b6306d

if (mode == "regression") {

[TSI-PTG]

Has there been any progress on this? I'm not successful with my attempts but maybe I'm doing something wrong.

library(tidymodels)
library(censored)
#> Loading required package: survival
library(stacks)
library(finetune)

tidymodels_prefer()

data(cancer)

lung <- lung %>%
    drop_na() %>%
    tibble() %>%
    mutate(surv = Surv(time, status))

lung_train <- lung %>% vfold_cv(strata = status, v = 5)

mod_glmnet <- proportional_hazards(penalty = tune(), mixture = tune()) %>%
    set_engine("glmnet") %>%
    set_mode("censored regression")


recipe_lung <- recipe(surv ~ ., data = lung) %>%
    update_role(c(inst, time, status), new_role = "ID")

workflow_lung <- workflow() %>%
    add_recipe(recipe_lung) %>%
    add_model(mod_glmnet)

tune_glmnet <- workflow_lung %>%
    tune_race_win_loss(
        lung_train,
        grid = 3,
        eval_time = 365,
        control = control_race(
            save_pred = TRUE,
            save_workflow = TRUE,
            allow_par = FALSE
        )
    )

stack_lung <- stacks() %>%
    add_candidates(tune_glmnet)
#> Error in `res[order(res$.row, res$.config), nms]`:
#> ! Can't subset columns that don't exist.
#> ✖ Columns `penalty` and `mixture` don't exist.
#> Backtrace:
#>      ▆
#>   1. ├─stacks() %>% add_candidates(tune_glmnet)
#>   2. ├─stacks::add_candidates(., tune_glmnet)
#>   3. ├─stacks:::add_candidates.tune_results(., tune_glmnet)
#>   4. │ └─... %>% .set_data_candidates(candidates, name, col_name)
#>   5. ├─stacks:::.set_data_candidates(., candidates, name, col_name)
#>   6. │ ├─... %>% dplyr::select(-.row)
#>   7. │ └─stacks:::collate_predictions(candidates)
#>   8. │   ├─tune::collect_predictions(x, summarize = TRUE) %>% ...
#>   9. │   ├─tune::collect_predictions(x, summarize = TRUE)
#>  10. │   ├─finetune:::collect_predictions.tune_race(x, summarize = TRUE)
#>  11. │   ├─base::NextMethod(summarize = summarize, parameters = parameters)
#>  12. │   └─tune:::collect_predictions.tune_results(...)
#>  13. │     └─tune:::average_predictions(x, parameters)
#>  14. │       └─tune:::surv_summarize(x, param_names, y_nms)
#>  15. │         ├─res[order(res$.row, res$.config), nms]
#>  16. │         └─tibble:::`[.tbl_df`(res, order(res$.row, res$.config), nms)
#>  17. │           └─tibble:::vectbl_as_col_location(...)
#>  18. │             ├─tibble:::subclass_col_index_errors(...)
#>  19. │             │ └─base::withCallingHandlers(...)
#>  20. │             └─vctrs::vec_as_location(j, n, names, missing = "error", call = call)
#>  21. ├─dplyr::select(., -.row)
#>  22. ├─tidyr::pivot_wider(...)
#>  23. ├─dplyr::mutate(., .config = process_.config(.config, df = ., name = col_name))
#>  24. ├─dplyr::select(...)
#>  25. ├─dplyr::mutate(., .config = if (".config" %in% names(.)) .config else NA_character_)
#>  26. ├─dplyr::ungroup(.)
#>  27. ├─dplyr::rename_with(., make.names, .cols = dplyr::starts_with(".pred"))
#>  28. └─vctrs (local) `<fn>`()
#>  29.   └─vctrs:::stop_subscript_oob(...)
#>  30.     └─vctrs:::stop_subscript(...)
#>  31.       └─rlang::abort(...)

^{Created on 2023-12-28 with reprex v2.0.2}

[EmilHvitfeldt]

As far as I know, there haven't been much progress yet. But we are still planning to make it happen!

[simonpcouch]

This will actually be a good bit more complicated than the kinds of changes in #157. The added functionality from accounting for evaluation times throughout stacks' internals is likely not worth the complexity, unfortunately.