Skip to content

mlr3extralearners 1.7.0

Latest

Choose a tag to compare

@be-marc be-marc released this 15 Sep 15:32
c512a29

New Features

  • surv.cv_glmnet and surv.glmnet: added the strata parameter for fitting stratified Cox models via glmnet::stratifySurv() (#396).

Breaking Changes

  • classif.mlp was renamed to classif.mlp_rsnns because its id clashed with the classif.mlp learner in mlr3torch.

Other

  • classif.catboost and regr.catboost now report $internal_valid_scores, which were previously always empty, for the eval_metric and the loss_function.

  • classif.catboost, regr.catboost, classif.lightgbm, regr.lightgbm, surv.xgboost.aft and surv.xgboost.cox gained a $best_valid_scores field, so msr("best_valid_score") can be used with them.

  • fix: classif.lightgbm and regr.lightgbm reported $internal_valid_scores for the last boosting iteration, although LightGBM predicts with model$best_iter. The scores are now taken from model$best_iter and therefore describe the model that is used for prediction.

  • New learners classif.tabfm and regr.tabfm interfacing the tabfm Python package, the tabular foundation model from Google Research.

  • classif.tabpfn and regr.tabpfn: added support for the feature types character, factor, and ordered, which are encoded as categorical features by tabpfn. The level order of ordered features is not preserved. Previously the features were converted to a numeric matrix, which ruled out categorical features even though tabpfn supports them.

  • New learners classif.bam and regr.bam fitting fast generalized additive models for large datasets with mgcv::bam() (#355).

  • New learner classif.opls fitting orthogonal partial least squares discriminant analysis with ropls::opls() from the Bioconductor package ropls (#268).

  • New learner regr.grf fitting a generalized random forest with grf::regression_forest(), supporting se predictions, observation weights, and missing feature values.

  • New learner regr.polynomial fitting a polynomial regression with stats::lm() and stats::poly() (#420).

  • classif.tabpfn and regr.tabpfn: updated the parameter sets to match tabpfn 8.1.0, adding auto_scale_n_estimators, keep_cache_on_device, n_preprocessing_jobs, differentiable_input, and show_progress_bar (plus eval_metric and tuning_config for classification), adding the "batched" option to fit_mode, and correcting the informational default of n_estimators to 8.

  • Fixed a partial argument-matching bug in survival glmnet learners where stype could be matched to predict argument s, causing s to be overwritten by stype = 1 or 2 and predictions to be over-regularized.