check tune::compute_metrics for survival models

[topepo]

Not sure if/where it is currently tested for this.

[topepo]

Looking at the code, I thought that we would need to add an eval_time argument (more on that below).

With no changes, the function does pretty much what it should (*)

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

lung_surv <- lung %>%
  dplyr::mutate(surv = Surv(time, status), .keep = "unused")

metrics <- metric_set(concordance_survival, brier_survival_integrated, brier_survival)

times <- c(2, 50, 100)

set.seed(2193)
tune_res <-
  proportional_hazards(penalty = tune(), engine = "glmnet") %>%
  tune_grid(
    surv ~ .,
    resamples = vfold_cv(lung_surv, 2),
    grid = tibble(penalty = c(0.001, 0.1)),
    control = control_grid(save_pred = TRUE),
    metrics = metrics,
    eval_time = times
  )

recomp <- compute_metrics(tune_res, metrics, summarize = TRUE)
original <- collect_metrics(tune_res)

all.equal(recomp, original)
#> [1] TRUE

stc_only <- compute_metrics(tune_res, metric_set(concordance_survival), summarize = TRUE)
stc_original <- collect_metrics(tune_res) %>% filter(.metric == "concordance_survival")

all.equal(stc_only, stc_original) # orginal has a .eval_time which is all NA
#> [1] "Names: 4 string mismatches"                                   
#> [2] "Length mismatch: comparison on first 7 components"            
#> [3] "Component 4: 'is.NA' value mismatch: 2 in current 0 in target"
#> [4] "Component 5: Mean relative difference: 0.6953599"             
#> [5] "Component 6: Mean relative difference: 58.79083"              
#> [6] "Component 7: Modes: character, numeric"                       
#> [7] "Component 7: target is character, current is numeric"
all.equal(stc_only, stc_original %>% select(-.eval_time))
#> [1] TRUE

^{Created on 2024-01-19 with reprex v2.0.2}

The only caveat is that the lack of an eval_time argument does not allow users to specify a subset of times. We could do that but the machinery to check that against the potential metrics change will add a substantial amount of code.

I'll look into where this is tested for survival models, but I think that we should leave the API as-is and see if there are requests for a subset of eval times. Also, I don't think that there are parity metrics for censored regression, so that is a moot point for that application.

[topepo]

We've settled on leaving compute_metrics() as-is. If we determine later that users would like to be able to filter evaluation times, we can make those changes.

We do need to make some unit tests for survival models so I'll keep this open until then.