mlr3proba 0.6.0

• Optimized surv.logloss and calib_alpha measures (bypassing distr6)
• Update/refine all measure docs (naming conventions from upcoming scoring rules paper) + doc templates
• fix very rare bugs in calib_alpha, surv.logloss and surv.graf (version with proper = FALSE)

mlr3proba 0.5.7

What's Changed

• Add breslow function for estimating the cumulative baseline hazard of proportional hazard models
• Add PipeOpBreslow to wrap a survival learner and generate distr predictions from lp predictions
• Add option breslow estimator option in distrcompositor
• Add extend_quantile to autoplot.PredictionSurv for type = "dcalib", which imputes NAs with the maximum observed survival time
• Fixes default in autoplot.PredictionSurv, now "calib"
• Update msr("surv.dcalib") default for truncate to Inf
• Add $reverse() method to TaskSurv, which returns the same task but with 1-status.
• Add reverse parameter to TaskSurv$kaplan() method, which calculates Kaplan-Meier on the censoring distribution of the task (1-status).
• Fix bottlenecks in Dcalib and RCLL

mlr3proba 0.5.3

What's Changed

• remove distr prediction from rpart survival learner by @bblodfon in #298
• generate_tasks.LearnerSurv respects feature types by @sebffischer in #249
• Update tic templates [ci-skip] by @pat-s in #300
• fix divide by zero bug in surv.cindex by @RaphaelS1 in #302
• ci: new workflows by @be-marc in #303
• fix ipcw bug by @RaphaelS1 in #304
• refactor: compability to mlr3viz 0.6.0 by @be-marc in #305
• Fix single prediction by @RaphaelS1 in #308
• fix na error in rcll by @RaphaelS1 in #310
• fix rcll and Uno's AUC bugs by @bblodfon in #314
• 0.5.0 by @RaphaelS1 in #316
• add regr.logloss by @RaphaelS1 in #317
• add probregr plot by @RaphaelS1 in #318
• fix rcll bug by @RaphaelS1 in #319
• fix test by @bblodfon in #320
• RCLL bug (1 observation in test set) by @bblodfon in #322
• Measures updates (RCLL fix range, change field values when ERV = TRUE) by @bblodfon in #324
• Optionally skip tests if 'rpart' unavailable by @MichaelChirico in #329
• Support for Arrdist prediction types by @bblodfon in #330
• Support arrdist by @RaphaelS1 in #331

New Contributors

• @MichaelChirico made their first contribution in #329

Full Changelog: v0.4.13...v0.5.3

mlr3proba 0.4.13

What's Changed

• fix AUCs by @RaphaelS1 in #268
• bugfix logloss by @RaphaelS1 in #272
• Fix Rcpp errors by @mllg in #270
• Prepare release by @mllg in #274
• add ERV by @RaphaelS1 in #276
• fix transpose bug by @RaphaelS1 in #277
• Closes survavg doesn't correctly output mixed distribution? #278 by @RaphaelS1 in #280
• chore: compatibility to mlr3 13.5 by @be-marc in #286
• bugfix by @RaphaelS1 in #289
• refactor: move plot functions from mlr3viz by @be-marc in #282
• Update tic templates [ci-skip] by @pat-s in #292
• Replace call to task() with S3 call + updated docs by @mllg in #293
• update doc by @bblodfon in #294

New Contributors

• @be-marc made their first contribution in #286
• @bblodfon made their first contribution in #294

Full Changelog: v0.4.7...v0.4.13

mlr3proba 0.4.7

mlr3proba 0.4.7

• Add right-censored log loss
• Fix bug in {rpart} where model was being discarded when set to be kept. Parameter model now called keep_model.

mlr3proba 0.4.6

• Patch for upstream breakages
• Add TaskSurv$kaplan method
• {survivalmodels} now imported (previously suggested)

mlr3proba 0.4.5

• Improved reduction from survival matrix predictions to ranking predictions
• Fixed cindex bug when all predictions equal
• Fix for valgrind

mlr3proba 0.4.4

• Minor change to how distributions are created to better support improper distributions
• Fixed bug in simsurv task that made it impossible to predict the target

mlr3proba 0.4.3

• Massive speed-up in distrcompositor PipeOp/pipeline
• More informative error given if $distr called for a learner that does not support this return type
• Fix massive bottleneck in scoring rule measures
• Add Density coercions as_task_dens and as_prediction_dens
• Measures now use parameter sets like learners. This streamlines the interface but unfortunately means ids can no longer be set dynamically.
• Add parameters t_max and p_max to Graf, Schmid and Integrated Log-loss as an alternative to times. t_max is equivalent to times = seq(t_max) and p_max is the proportion of censoring to integrate up to in the dataset.
• Fix bug in Rcpp code that was causing erroneous values for calculating the cindex in datasets greater than 20,000 observations.

mlr3proba 0.4.2

• Patch for linux

mlr3proba 0.4.1

• Remove mlr3extralearners from Suggests
• Add response to as_prediction_surv
• Now exported a couple cpp functions and assert_surv
• mlr3 is now in Depends not imports
• distr predictions are now internally stored as matrices to significantly reduce prediction object sizes
• Tasks now support strata property

mlr3proba 0.4.0

• Deprecated measures from 0.2.0 have now been deleted.
• IPCW measures such as surv.graf, surv.schmid, and surv.intlogloss now allow training data to be passed to the score function with task and train_set to allow the censoring distribution to be estimated on the training data. This is automatically applied for resample and benchmark results.
• IPCW measures such as surv.graf, surv.schmid, and surv.intlogloss now include a parameter proper to determine what weighting scheme should be applied by the estimated censoring distribution, The current method (Graf, 1999) proper = FALSE, weights observations either by their event time or 'current' time depending if they're dead or not, the new method proper = TRUE weights observations by event time. The proper = TRUE method is strictly proper when censoring and survival times are independent and G is estimated on large enough data. The proper = FALSE method is never proper. The default is currently proper = FALSE to enable backward compatibility, this will be changed to proper = TRUE in v0.6.0.
• The rm_cens parameter in surv.logloss has been deprecated in favour of IPCW. rm_cens will be removed in v0.6.0. If rm_cens or IPCW are TRUE then censored observations are removed and the score is weighted by an estimate of the censoring distribution at individual event times. Otherwise if rm_cens and IPCW are FALSE then no deletion or weighting takes place. The IPCW = TRUE method is strictly proper when censoring and survival times are independent and G is estimated on large enough data. The ipcw = FALSE method is never proper.
• Add surv.dcalib for the D-Calibration measure from Haider et al. (2020).