Repository navigation
New Features
surv.cv_glmnetandsurv.glmnet: added thestrataparameter for fitting stratified Cox models viaglmnet::stratifySurv()(#396).
Breaking Changes
classif.mlpwas renamed toclassif.mlp_rsnnsbecause its id clashed with theclassif.mlplearner inmlr3torch.
Other
-
classif.catboostandregr.catboostnow report$internal_valid_scores, which were previously always empty, for theeval_metricand theloss_function. -
classif.catboost,regr.catboost,classif.lightgbm,regr.lightgbm,surv.xgboost.aftandsurv.xgboost.coxgained a$best_valid_scoresfield, somsr("best_valid_score")can be used with them. -
fix:
classif.lightgbmandregr.lightgbmreported$internal_valid_scoresfor the last boosting iteration, although LightGBM predicts withmodel$best_iter. The scores are now taken frommodel$best_iterand therefore describe the model that is used for prediction. -
New learners
classif.tabfmandregr.tabfminterfacing thetabfmPython package, the tabular foundation model from Google Research. -
classif.tabpfnandregr.tabpfn: added support for the feature typescharacter,factor, andordered, which are encoded as categorical features bytabpfn. The level order oforderedfeatures is not preserved. Previously the features were converted to a numeric matrix, which ruled out categorical features even thoughtabpfnsupports them. -
New learners
classif.bamandregr.bamfitting fast generalized additive models for large datasets withmgcv::bam()(#355). -
New learner
classif.oplsfitting orthogonal partial least squares discriminant analysis withropls::opls()from the Bioconductor packageropls(#268). -
New learner
regr.grffitting a generalized random forest withgrf::regression_forest(), supportingsepredictions, observation weights, and missing feature values. -
New learner
regr.polynomialfitting a polynomial regression withstats::lm()andstats::poly()(#420). -
classif.tabpfnandregr.tabpfn: updated the parameter sets to matchtabpfn8.1.0, addingauto_scale_n_estimators,keep_cache_on_device,n_preprocessing_jobs,differentiable_input, andshow_progress_bar(pluseval_metricandtuning_configfor classification), adding the"batched"option tofit_mode, and correcting the informational default ofn_estimatorsto8. -
Fixed a partial argument-matching bug in survival
glmnetlearners wherestypecould be matched to predict arguments, causingsto be overwritten bystype = 1or2and predictions to be over-regularized.