OOB with ranger: Error during tuning for TTE endpoints

[MasterLuke84]

Hi Raphael,

I hope this is the right repo for this issue.

When using oob_error as performance measure for TTE endpoint with the ranger engine this error occurs:

It works with regression/classification endpoints.

Example

library(mlr3)
library(paradox)
library(mlr3proba)
library(survival)

data("cancer")
data <- cancer %>% na.omit()

params <- ps(
  mtry =          p_int(lower = 1, upper = 3), 
  min.node.size = p_int(lower = 1, upper = 2), 
  num.trees =     p_int(lower = 50, upper = 100))

grid <- generate_design_lhs(params, n = 3)

mlr3tuning::tune(method       = "design_points", 
                 learner      = lrn("surv.ranger"), 
                 task         = as_task_surv(data, time = "time", event = "status",  id = "1111"), 
                 resampling   = mlr3::rsmp("cv", folds = 4), 
                 measures     = mlr3::msr("oob_error"), 
                 design       = grid$data, 
                 store_models = TRUE,
                 search_space = params)

Best regards

[RaphaelS1]

Thanks. It's breaking because of incompatible predict types, I might have to implement surv.oob_error to fix this.

[MasterLuke84]

At some point it has already worked. If you run the example with following package versions, no issues occur.

mlr3 0.12.0
mlr3learners 0.5.0
mlr3proba 0.4.1
mlr3tuning 0.9.0
mlr3measures 0.4.0

Maybe this might help to narrow down the issue.

Thanks in advance and enjoy the weekend!

[bblodfon]

@RaphaelS1 you can close this (fixed in https://github.com/mlr-org/mlr3/issues/816)