A Large-Scale Neutral Comparison Study of Survival Models on Low-Dimensional Data

L. Burk, J. Zobolas, B. Bischl, A. Bender, M. N. Wright, and R. Sonabend, “A Large-Scale Neutral Comparison Study of Survival Models on Low-Dimensional Data.” arXiv, Jun. 06, 2024. doi: 10.48550/arXiv.2406.04098.

Setup

The benchmark is conducted using R and the mlr3 framework. The following files are necessary to set up the benchmark:

• config.yml contains configuration defaults for different settings (pretests, debugging, etc.)
  • It is loaded using config::get() in the main scripts.
• batchmark.R is the main setup script. It creates a benchmark registry (using batchtools and mlr3batchmark).
  • Resulting compute jobs can be submitted with batchtools::submitJobs() depending on available resources.
  • submit.R contains job management steps specific to the HPC environment used for the benchmark.
  • To actually run the benchmark, you need to adjust the submit.R script to your environment! Please refer to the batchtools documentation for more information.
• Datasets are loaded from ./datasets in .rds format, where they are also stored in CSV and arff formats.
  • Code for retrieval and minimal preprocessing is in import_datasets.R
  • In batchmark.R, the datasets are loaded and converted into {mlr3} TaskSurv objects.
• resamplings contains resampling fold information for each dataset in CSV form to ensure reproducibility.
  • These files are automatically generated when running batchmark.R, where resampling is done using a set RNG seed.
• helpers.R contains helper functions for benchmark setup and later analysis.
• site/ contains a quarto site with all results.

Reproducibility

• renv and renv.lock contain renv project information.
  • renv::restore() needs to be used to restore the project environment.
  • Due to limitations on the HPC environment used for the benchmark, R version 4.2.2 is expected. (you may want to use [rig](renv for R version management)
  • Additionally, the PMCMRplus is required by mlr3benchmark, but was not installable in the HPC environment, so it is not included by renv and you may need to install it manually from CRAN: install.packages("PMCMRplus").
• produce_paper_plots.R contains code to reproduce the plots used in the paper and aims to be as self-contained as possible, but loads helper functions from helpers.R for de-duplication and readability.
  • The output path is ./results_paper by default.

Please note that due to the large file sizes of the BenchmarkResult (bmr) objects produced by the aggregation of the batchtools registry, this repository only contains the processed result files (./results/registry_beartooth/) required to produce the main results of the paper.

Results

Results are available online at projects.lukasburk.de/survival_benchmark/

The site is generated from the quarto site in ./site/.

Datasets

The datasets used in the benchmark are stored after minor modifications in ./datasets/ and are also uploaded to OpenML. The dataset's names, source package, and OpenML dataset IDs are stored in ./dataset_table.[csv|rds].

Here is a short example on how to download the datasets from OpenML using {mlr3oml}:

# Get dataset from openml
library(mlr3oml)
library(mlr3proba) # To create survival tasks via as_task_surv

# The 'qs' package is required for caching the downloaded data
if (requireNamespace("qs", quietly = TRUE)) {
  options(mlr3oml.cache = TRUE)
}

# Get the table of datasets & their OpenML IDs
dataset_tbl = readRDS(here::here("dataset_table.rds"))
head(dataset_tbl[, c("dataset", "dataset_id")])

   dataset dataset_id
1     gbsg      46131
2 metabric      46142
3  support      46144
4   colrec      46145
5    rdata      46146
6  aids.id      46130

# Get an individual dataset in the OMLData class
colrec_odt = mlr3oml::odt(46145)

# Convert the OMLData object to a TaskSurv object in a loop, creating a list of mlr3 TaskSurv objects
task_list = lapply(dataset_tbl$dataset_id, function(id) {
  dat = mlr3oml::odt(id)

  task = mlr3proba::as_task_surv(mlr3::as_data_backend(dat), target = "time", event = "status", id = dat$name)
  task$set_col_roles("status", add_to = "stratum")
  Sys.sleep(0.1) # Small timeout to not hammer the OML server
  task
})

task_list[[1]]

<TaskSurv:gbsg> (2232 x 9)
* Target: time, status
* Properties: strata
* Features (7):
  - int (7): x0, x1, x2, x3, x4, x5, x6
* Strata: status