survival learner issues in via mlr3

[bblodfon]

Hi Byron,

Just tried to use aorsf with survival data via mlr3extralearners and got this error:

library(mlr3extralearners)
library(mlr3pipelines)
library(mlr3proba)
#> Loading required package: mlr3

task = tsk('lung')
pre = po('encode', method = 'treatment') %>>%
  po('imputelearner', lrn('regr.rpart'))
task = pre$train(task)[[1]]
task
#> <TaskSurv:lung> (228 x 10): Lung Cancer
#> * Target: time, status
#> * Properties: -
#> * Features (8):
#>   - int (7): age, inst, meal.cal, pat.karno, ph.ecog, ph.karno, wt.loss
#>   - dbl (1): sex

aorsf = lrn('surv.aorsf', control_type = 'fast',
  oobag_pred_type = 'surv', importance = 'anova',
  attach_data = TRUE)

aorsf$train(task)
#> Error: some variables have unsupported type:
#>  <status> has type <logical>
#> supported types are numeric, integer, units, factor, and Surv
aorsf$errors
#> character(0)

^{Created on 2024-04-11 with reprex v2.0.2}

Maybe you haven't updated it with recent changes? I see for example that the oobag_pred_type default is risk now, not surv.

[bblodfon]

Btw is the mort prediction type the same as in RSFs? Eg adding the cumulative hazard for all time points for each observation? Something like:

cumH = -log(surv_matrix) # surv_mat is (obs x times)
cumH[cumH == Inf] <- 0
rowSums(cumH)

PS: I think it is so after looking a bit at the doc and running some examples, but good to verify it form you!

I am using the oobag error for wrapped-based feature selection with aorsf via mlr3extralearners and I think that using mort as a default oob_pred_type might be the best option as it is time-independent, rather than calculating the C-index on the median survival time that is the default now. What do you think?

[bcjaeger]

Hi John, great questions!

1. The error you received is due to status being a logical variable. I wrote that feature a long time ago and it seems kind of unnecessary now. Would it be helpful if I made a small patch in aorsf so that logical variables are converted to numeric 0/1 integer-type variables?
2. You caught me. I made the change to default risk when growing trees. It always bugged me that the default oobag prediction type in aorsf was survival while the default predict type was risk. I am sorry if this caused backward compatibility issues. =[
3. Yes, mort is intended to match the mortality prediction in RSF. I agree with using mort to guide variable selection. In fact, mort is automatically used in aorsf's built-in variable selection function: orsf_vs

[bblodfon]

1. Yes please! It make sense with status, only a few datasets have TRUE/FALSE instead of an integer 0 or 1 and if you want to go multi event in the future, checking/converting to integer is the way to go.
2. Haha yes, let's just make sure everything is updated when you have the time in mlr3extralearners, I can double-check the PR if you want. Including the response = rmst prediction could be also included as an extra that would be nice.
3. Super

I will be heavily using aorsf soon and I was glad to find other things to be solved. There was some C++ issue when running with extremely large datasets (~30K features) if I find it again, I will let you know!

[bcjaeger]

I will be heavily using aorsf soon

This absolutely made my day. Thank you!!

I will get to work on the patch and the addition of rmst in mlr3extralearners soon (sorry I have been slow on the rmst PR - it is on my short list of priorities now).

[bblodfon]

Hi @bcjaeger! Would you have time to quickly look and update the PR in mlr3extralearners?

[bblodfon]

I updated the tsk("lung") of mlr3proba to encode status as 0/1 so the original issue is no longer present.