survdistr.mlr3cmprsk version 0.0.5.crs parameters.fastai and botorch tests on Windows and macOS where the Python
backends crash or time out.tabpfn tests until token work reliable again.blockForest tests on macOS where SE predictions fail sanity checks.h2o.glm classification tests on Windows due to Java NullPointerException.GPfit tests on Windows where they crash under R-devel.classif.aorsf sanity autotest due to inconsistent tie-breaking across
predict types.surv.flexreg sanity autotest on Windows due to initial parameter
estimation failure.LearnerCompRisksCoxboostLearnerRegrGPfitLearnerClassifMLPLearnerClassifSaeDNNLearnerClassifPlsdaCaretLearnerSurvDNNLearnerRegrH2ORandomForestLearnerRegrH2OGLMLearnerClassifH2OGLMLearnerClassifH2OGBMLearnerClassifH2ORandomForestLearnerClassifH2ODeeplearningLearnerRegrH2OGBMLearnerRegrH2ODeeplearningLearnerClassifLvq1LearnerRegrBotorchFullyBayesianLearnerRegrBotorchSingleTaskGP and LearnerRegrBotorchMixedSingleTaskGP.LearnerSurvAkritas and LearnerSurvParametric to the attic.
See https://github.com/mlr-org/mlr3extralearners/issues/549.Extending vignette to incorporate information about skipping tests and considerations for testing Python learnerssurvdistr is now on Suggests (used for constant interpolation of the Kaplan-Meier predictions of the partykit survival learners)mlr3proba (0.8.8), pls and xgboost to the most recent CRAN versionsLearnerSurvGamCoxLearnerSurvFlexRegLearnerSurvNCVsurvLearnerRegrRRFLearnerRegrPcrLearnerRegrPlsrLearnerRegrLaGPLearnerRegrFrbsLearnerRegrBcartLearnerRegrBgpLearnerRegrBgpllmLearnerRegrBlmLearnerRegrBtgpLearnerRegrBtgpllmLearnerRegrBtlmLearnerRegrNCVRegLearnerClassifDbnDNNLearnerClassifNNTrainLearnerClassifSparseLDALearnerClassifNCVreglrn("surv.flexible") (LearnerSurvFlexible) was renamed to lrn("surv.flexsurvspline") (LearnerSurvFlexSpline) to properly reflect the wrapped train function (Royston/Parmar spline model).CoxBoost is now on CRAN, so we removed it from Remoteslrn("surv.flexsurvspline") predicts linear predictors using predict.flexsurvreg(). We were doing manually the same exact prediction, so no functionality was changed.xgboost 3.1.2.1 (survival learners)regr.lmer/glmer learnersrandomForestSRC 3.5.0 (use.uno parameter)data.table::fifelse (@m-muecke)formula and anc params to surv.flexible learner, as well as response predict type (mean survival time).regr.gamboost regression predictions (#498).New Learners:
LearnerCompRisksRandomForestSRCLearnerSurvBlockForestLearner{Classif,Regr,Surv}BlockForestLearner{Classif,Regr}ExhaustiveSearchLearnerClassifFastaiLearner{Classif,Regr}PenalizedLearner{Classif,Regr}BstLearnerClassifAdabagLearnerClassifAdaBoostingLearner{Classif,Regr}EvtreeLearnerClassifKnnLearnerClassifRotationForestLearnerRegrCrsLearnerClassifStepPlrLearnerClassifMdaLearnerClassifRfernsLearnerClassifNeuralnetLearnerRegrBrnnLearnerRegrBotorchSingleTaskGPLearnerRegrBotorchMixedSingleTaskGPAdd new control_custom_fun parameter in surv.aorsf
New function learner_is_runnable() to check whether the
required packages to train a learner are available.
Added selected_features property to RandomForestSRC learners (prediction doesn't work if vars.used = 'all.trees')
discrete parameter from surv.parametric, so that it is impossible to return distr6::VectorDistribution survival predictions (softly deprecated in [email protected])mlr3proba soon (see v0.8.2 or later).create_learner() generator was removed, because it was hard to maintain and boilerplate code in the age of LLMs is easy enough to write.discrete parameter from surv.parametric, so that it is impossible to return distr6::VectorDistribution
survival predictions (softly deprecated in [email protected])classif.lightgbm now works with encapsulation with multiclass taskslrn and lrns, which should anyway
be available to the user as the package depends on mlr3, where these
functions are defined.randomPlantedForest was removed, because there is currently no way to
save the model.survivalmodels were removed, because
they also cannot be saved and because the upstream package is archived.withrmlr3proba is now an import and no longer a suggested package.mlr3cmprsk is added as an import.set.seed() in the tests and instead uses withr::local_seed()
This means the auto tests will be stochastic like they should be.distr6 dependency is removed. partykit survival learners use constant
interpolation of the predicted Kaplan-Meier curves via survdistr::vec_interp()New Features:
regr|classif.mgcv, regr.glm and regr.lmer.LearnerRegrQGam and LearnerRegrMQGam.LearnerClassifTabPFN and LearnerRegrTabPFN.surv.xgboost.cox.LearnerClassifKnn.Bugfixes:
crank as main prediction type (and it is always returned) #331DESCRIPTION filegridify_times() function to coarse time pointssurv.parametric and surv.akritas use of ntime argumentsurv.parametric is now used by default with discrete = TRUE (no survival learner returns now distr6 vectorized distribution by default)mlr3 (version 0.21.0)learner$importance()n_thread = 1 for surv.aorsf and use unique event time points for predicted S(t)selected_features() for surv.penalizedsurv.prioritylasso learner + add distr predictions via Breslow #344gamma.mu parameter was split to gamma and mu to enable easier tuning (surv.svm learner)aorsf learnerresample() or benchmark() (#353)$model slot of lrn("classif.abess") now contains the model of
the upstream package again.lrn("surv.xgboost.aft")
and lrn("surv.xgboost.cox").case.depth parameter to rfsrc learners.mlr3 is now in Depends instead of Imports.lrn("surv.xgboost") was now removed.
Use lrn("surv.xgboost.cox") or lrn("surv.xgboost.aft") instead.surv.xgboost.cox and surv.xgboost.aft separate survival learners. distr prediction on the cox xgboost learner is now estimated via Breslow by default and aft xgboost has now in addition a response prediction (survival time)surv.parametric code to survivalmodels, changed type parameter to form to avoid conflict with survivalmodels's default parameter listVectorDistributions from partykit and flexsurv survival learners with survival matrices (Matdist) (thanks to @bblodfon)discrete parameter in surv.parametric learner to return Matdist survival predictionsselected_features() to CoxBoost survival learners (thanks to @bblodfon)surv.ranger now receives parameters during $predict() (thanks to @jemus42)surv.bart was added (thanks to @bblodfon)lrn("surv.aorsf") were updated (thanks to @bcjaeger)distr predict type to the surv.cv_glmnet and surv.glmnet
learners (thanks to @bblodfon)I and F params from IBk learner are too interdependent (I can only be TRUE when F is FALSE and vice versa).
Combined them into one factor param weight that has two levels -- I and F.U must be FALSE for S to be tunable in J48 learner.perf.type to rfsrc learners"multiclass" property from lrn("classif.gbm"), as this feature is broken."factor" to gam learnersmin.bucket to rangernthreads to dbarts learners; set verbose to FALSE by default (thanks to @ck37)regr.abess and classif.abess (thanks to @bbayukari)LearnerClassifImbalancedRandomForestSRC (thanks to
@HarutyunyanLiana)LearnerClassifPriorityLasso, LearnerRegrPriorityLasso, LearnerSurvPriorityLasso (thanks to
@HarutyunyanLiana)LearnerClassifGlmer (https://github.com/mlr-org/mlr3extralearners/issues/243)nei and ncv.thread that were added to mgcv::gam() in
version 1.8.41"weights" to LearnerClassifGlmer and
LearnerRegrLmerlightgbm uses the param_vals stored in the state for hotstartingstate$data_prototype to get ordering of features via
ordered_features() like in mlr3learners and therefore obviate the need to
store feature_names in the statemlr_learners when unloading
mlr3extralearners"integer" to classif.randomForest"logical" to {classif, regr}.randomForestlist_mlr3learners() function. Now slower but correct.keep_data instead of keep.dataLearnerSurvAorsf with key surv.aorsf. See https://github.com/bcjaeger/aorsf for more details on aorsfcreate_learner and the learner template.Fix bug in C50 learner: Weights were not passed correctly
Remove kerdiest Learner because it is not being maintained on CRAN anymore
Fix bugs in learners lmer and J48
Remove predict type proba from J48
Delay loading of mlr3proba learners
lightgbm:
Docs: Renamed section "Custom mlr3 defaults" to "Parameter Changes"
Added labels to learners
Remove extraTrees because it is no longer on CRAN and GH version has errors
Remove sketch_eps parameter from xgboost because it is no longer listed in the docs
regr.lmerchore: add new parameters for kde and rfsrc
temporarily disable feat_all test for obliqeRSF (failed in $score() stage, because issue only happened in CI and could not be reproduced
early_stopping_split for lightgbm learnersdens.plugcreate_learnercategorical_features in {lightgbm} learnerslightgbm updatescreate_learnerignored_features and one_hot_max_size to classif.catboostregr.rvm and classif.lssvmrandomForestSRC::rfsrc(),
partykit::cforest() and obliqueRSF::ORSF() to conveniently tune
hyperparameters whose upper limit depends on data dimensions.regr.gaussprregr.gausspr and classif.gausspr from kernlab::gaussprinstall_catboost to make installation from catboost simplerbase parameter of {bart} learnersLearnerRegrCubist and LearnerRegrMarsnnet learners to mlr3learnersrfsrc learners to 1LearnerRegrGam and LearnerClassifGam with keys regr.gam and classif.gam from package mgcv.surv.coxboost now uses the GitHub version instead of CRAN (archived)regr.glmboostsurv.svm now supports all feature typesLearnerRegrLightGBM and LearnerClassifLightGBM with keys regr.lightgbm and classif.lightgbm respectively. Copied from mlr3learners.lightgbmLearnerRegrLiblineaRX and LearnerClassifLiblineaRX deprecated in favour of only two learners (LearnerRegrLiblineaR and LearnerClassLiblineaR) with added hyper-parameters. Deprecated learners will be removed in v0.3.0.classif.nnet will be removed in v0.4.0.liblinearX will be removed in v0.4.0.dist = "logistic" has been removed from surv.parametric as it is unclear what this was previously predicting.type = "tobit" for dist = "gaussian" so predictions can correspond with survival::survreg.LearnerRegrGlm with the unique key regr.glm from package stats, which allows users to change the family hyperparameter when fitting generalized linear regression models.keeptrees parameter from classif.bart as this is forced internallyclassif.bartclassif.earth and regr.earthse predict type to regr.earthregr.knn and classif.knnmlr3proba moved to Suggestsinstall_learners now additionally installs required mlr3 packagessurv.parametric occurring if feature names are switched between training and predictingclassif.nnet, in the future please load from mlr3learnerscrank and distr computation of all survival learnerssurv learners that were reversing the order of crank, see this issue for full details: https://github.com/mlr-org/mlr3proba/issues/165response is no longer returned by surv.mboost, surv.blackboost, surv.glmboost, surv.gamboost or surv.parametricsurv.parametric with ph formsurvivalmodelslearners which weren't returning distrsurv.coxboost and surv.coxboost_cv can now only handle integer and numeric feature types, previous automated internal coercions were inconsistent with mlr3 design.