NEWS

---

mlr3resampling 2026.5.19 (2026-05-19)

• remove ResamplingSameOtherCV and ResamplingVariableSizeTrainCV.

mlr3resampling 2026.5.15

• remove animint2 from Suggests, remove Older resamplers vignette.
• figure and text size fixes in vignettes.

mlr3resampling 2026.5.14

• pvalue()$pvalues table no longer has diff_mean column (was negative of mean_diff column, and potentially confusing).
• pvalue_downsample() stops if downsample results are missing, and messages if there are several combinations of task_id, algorithm, test.subset.
• pvalue_downsample()$caption includes task_id.
• plot.score() sorts y axis by Train_subsets and n.train.groups.
• add SOAKED vignette with two Tasks of simulated data (iid_easy has sample size effect, different has subset effect).

mlr3resampling 2026.5.10

• WasikowskiLinearMemory changed to WasikowskiLimitedMemory.

mlr3resampling 2026.5.1

• New subset_group_stratum vignette, with analysis of counts per fold, using column roles group, stratum, subset.
• ResamplingSameOtherSizesCV$instantiate() gains param group_stratum_algo, which is used to determine the fold assignment algorithm for Tasks with groups and strata (even if each group is in only one stratum). Default "RSS" uses novel C++ code, linear time heuristic for minimizing residual sum of squares (RSS), between actual and ideal counts in each stratum/fold. The previous method can be used via "WasikowskiLinearMemory" (named after author of Kaggle post), and is slightly slower and less accurate (in terms of RSS).

mlr3resampling 2026.4.26

• ResamplingSameOtherSizesCV$instantiate() uses new C++ code, adapted from scikit-learn StratifiedGroupKFold, to assign folds, when there are groups in several strata.

mlr3resampling 2026.4.21

• ResamplingSameOtherSizesCV$instantiate() stops if group in several strata.

mlr3resampling 2026.4.15

• proj_grid() gains argument train_seed, default=1L for reproducibility, use NA_integer_ for old behavior (seed not set before training).

mlr3resampling 2026.4.13

• ResamplingSameOtherSizesCV with sizes=0 now means 5 train/test splits per test subset, instead of 6. The intermediate size was removed because it is not used in pvalue_downsample().

mlr3resampling 2026.4.9

• proj_grid() error for train/test set with missing class.
• new proj vignette.
• proj_compute_all() gains LAPPLY arg.
• proj_test() gains verbose, LAPPLY args.
• ResamplingSameOtherSizesCV respects new column role=fold, for reproducible splits in benchmarks. If fold role is set, it should also be used in stratum role.

mlr3resampling 2026.3.16

• new pvalue_downsample() with plot() method for SOAK downsampling results.
• new internal pvalue_compute() and pvalue_prep() for shared p-value computation logic used by pvalue() and pvalue_downsample().

mlr3resampling 2026.3.10

• fixed integer overflow in large Tasks (PR#59).
• proj_compute_all() uses future.seed=NULL.
• pbdMPI moved from Imports to Suggests for installation on Mac.
• proj_submit() stops if pbdMPI not available.

mlr3resampling 2026.2.24 (2026-02-26)

• proj_compute_all() uses future_lapply() if available.

mlr3resampling 2026.2.20

• added inst/CITATION.

mlr3resampling 2025.11.19 (2025-11-20)

• plot() yields consistent Y axes (other on top) for score() and pvalue().

mlr3resampling 2025.9.24

• Newer resamplers vignette reproducibility section.

mlr3resampling 2025.9.22

• Older resamplers vignette uses HTML grid layout.
• score() uses setattr() instead of class()<- to set "score" class of data table output.

mlr3resampling 2025.7.30

• proj_grid() outputs tasks and resamplings sub-directories, instead of saving these data to the grid.rds file.
• proj_compute_mpi() launches a SLURM MPI job in which each task computes a variable number of results.
• proj_test() for running one CV iteration, on down-sampled Tasks, using edited Learners.
• proj_fread() for reading all CSV results.
• New Learners AutoTunerTorch_epochs, LearnerRegrCVGlmnetSave, LearnerClassifCVGlmnetSave, with special methods for proj_* functions: save_learner and edit_learner.

mlr3resampling 2025.6.23 (2025-06-23)

• proj_* functions for SLURM load balancing (PR#32).

mlr3resampling 2025.5.21

• ResamplingBase inherits from mlr3::Resampling (PR#31, thanks Sebastian Fischer).

mlr3resampling 2025.5.17

• ResamplingBase defines instantiate() which now saves task hash (PR#26, thanks Marc Becker); sub-classes define get_instance().

mlr3resampling 2025.3.30 (2025-04-09)

• new plot() method for score().
• pvalue() with plot() method.
• ResamplingSameOtherSizesCV gains subsets parameter which specifies which subsets to use for training: S, O, A, SO, SA, SOA (S=same, O=other, A=All).

mlr3resampling 2024.10.28

• .onLoad, add mlr3resampling to mlr3_reflections$loaded_packages, thanks @be-marc for the idea, https://github.com/tdhock/mlr3resampling/issues/17

mlr3resampling 2024.10.22

• update docs, group -> subset.

mlr3resampling 2024.9.6 (2024-09-11)

• .onLoad, add subset to mlr3_reflections$task_col_roles for classif and regr.

mlr3resampling 2024.7.14

• fix typo newere in vignette.

mlr3resampling 2024.7.7 (2024-07-12)

• use purl=FALSE in vignettes to avoid CRAN check error.

mlr3resampling 2024.7.3 (2024-07-05)

• use mlr3tuning::tnr("grid_search") instead of TunerGridSearch in Newer_resamplers vignette (same functionality, new name is TunerBatchGridSearch).
• fix Rd files Task -> mlr3:Task, rm ResamplingHoldout etc.

mlr3resampling 2024.5.14

• mlr3misc::leanify_package() and add ResamplingSameOtherSizesCV to mlr_resamplings as same_other_sizes_cv, thanks @sebffischer.

mlr3resampling 2024.5.13

• rename ResamplingSameOtherSizes.Rmd to Newer_resamplers.Rmd to avoid error on windows, file name conversion problem -- name too long?

mlr3resampling 2024.5.2

• new section AZtrees in ResamplingSameOtherSizes vignette, and updated data set (factor columns instead of int/char).
• to save disk space, consolidated vignettes.

mlr3resampling 2024.4.14 (2024-04-16)

• ResamplingSameOtherSizesCV and ResamplingSameOtherCV now recognize subset (instead of previous group_generalization/group) as the column role for designating subsets to train/test on.
• ResamplingSameOtherSizesCV now recognizes group (instead of previous group_atomic) as the ID of grouped observations that should not be separated when splitting.

mlr3resampling 2024.3.6

• new ResamplingSameOtherSizesCV.

mlr3resampling 2024.1.23 (2024-02-01)

• ResamplingVariableSizeTrainCV outputs train sets which respect strata.

mlr3resampling 2024.1.8 (2024-01-09)

• Rename Simulations vignette to ResamplingSameOtherCV.
• New ResamplingVariableSizeTrainCV class and vignette.
• use setDTthreads(1L) before tests to avoid CRAN NOTE (Running R code in ‘testthat.R’ had CPU time 2.8 times elapsed time).

mlr3resampling 2023.12.23

• To get data set names in Simulations vignette, use task data names instead of learner$state$data_prototype.

mlr3resampling 2023.12.20 (2023-12-21)

• First CRAN release.
• Link to blog in DESCRIPTION.

mlr3resampling 2023.12.19

• use setDTthreads(1) in vignette to avoid CRAN NOTE (Re-building vignettes had CPU time 3.2 times elapsed time)

mlr3resampling 2023.12.18

• do not future::plan multisession, to avoid CRAN NOTE (checking for detritus in the temp directory).

mlr3resampling 2023.12.13

• Error for conflicting group names.
• instance has new elements: data tables for visualization.

mlr3resampling 2023.12.12

• Examples vignette, first tests, score().

mlr3resampling 2023.12.11

• minimal docs, tests, etc.

mlr3resampling 2023.12.8

• first version, R code forked from https://tdhock.github.io/blog/2023/R-gen-new-subsets/