NEWS.md

nadir 1.0.0

This release marks {nadir}'s readiness to submit to rOpenSci for peer review, and incorporates everything described under 0.0.2 below (which was a development version and was never released to CRAN).

Breaking changes

• Learner backend packages (e.g., glmnet, ranger, xgboost, lightgbm, dbarts, e1071, earth, gbm, hal9001, kernlab, kknn, lme4, mgcv, nnet, randomForest, rpart, VGAM, and others) have been moved from Imports to Suggests, in order to reduce {nadir}'s installation footprint. Users only need to install the backends for the learners they actually use. A new internal require_backend() check is called throughout, so attempting to use a learner whose backend is not installed produces an informative error naming the package to install.

Documentation and vignettes

• New Glossary article defining the technical terms used throughout the package documentation (cross-fitting, out-of-fold predictions, metalearning, and so on).
• New Translation Guide article for users arriving from {SuperLearner} or {sl3}, which supersedes and replaces the previous comparison article.
• Expanded the Guidance for Developers article and FAQs.

Internal changes

• Completed pkgcheck::pkgcheck() compatibility in preparation for rOpenSci submission.
• Tests for Suggests-backed learners are now skipped gracefully when the corresponding backend package is not installed.

nadir 0.0.2

New features

• crossfit_super_learner() is a newly available method for cross-fitting super learners; cv_super_learner() has been reworked to use crossfit_super_learner() internally.
• super_learner() output gains $oof_predictions() and $oof_predict_modified() for accessing and working with out-of-fold predictions, alongside $oof_predict() and $oof_predict_fold().
• New train_on_whole_dataset argument to super_learner() (default TRUE) allows users to skip refitting learners on the full dataset after cross-validation.
• add_stratification() added for stratified learner construction.
• truncate_lnr() added for bounding a learner's predictions between user-specified minimum and maximum values.
• Learners that error or warn during fitting now have their messages captured and stored for user inspection rather than cluttering the console; a new vignette covers timing learners via warnings.
• Support for multi-predictors: learners that return lists of multiple predictors, via as_multi_predictor().
• Formulas may now be passed as character strings (including a single string recycled across learners), which are handled via as.formula().
• Fit learners are now returned as part of the super learner object ($fit_learners).
• $predict() on super learned models improved, including support for newdata that omits the outcome column.

New learners

• lnr_lightgbm() (LightGBM)
• lnr_bart() (Bayesian Additive Regression Trees), with careful handling of prediction closures when categorical variables are present
• lnr_knn() and lnr_knn_binary() (k-nearest neighbors)
• lnr_svm() and lnr_svm_binary() (support vector machines)
• lnr_cvglmnet() (cross-validated glmnet)

New S3 methods

• nadir_sl_model and nadir_crossfit_sl objects gain print(), plot(), summary(), coef(), nobs(), formula(), fitted(), and residuals() methods.
• plot() method added for nadir_cv_sl objects.
• fitted() and residuals() are returned in the row-order of the input data via an internal .sl_rowid, with tests ensuring correctness on out-of-order and shuffled inputs.

Bug fixes

• Fixed predictions on data whose columns are out of order relative to the training data.
• Fixed an obs_weights bug in weight determination.
• lnr_glmnet() (and lnr_glmnet_grid()) now work even when some factor levels contain zero observations; rare-factor handling also added to lnr_hal() and lnr_hal_grid().
• Fixed a bug in lnr_gbm() when a weights = NULL argument was passed.
• Fixed bugs in lnr_xgboost() and lnr_lightgbm(); updated lnr_xgboost() so binary learners work again.
• Fixed n_folds handling in cv_super_learner().
• cv_origami_schema() now errors informatively on non-discrete cluster_ids / strata_ids.
• The negative log loss for binary outcomes now truncates predicted probabilities (following the same truncation scheme as sl3) to avoid -log(0) = Inf.
• Weight determination now checks edge cases: a single learner, and all-zero model weights (falling back to equal weights), with informative errors added throughout determine_weights_*().
• Outcome variables are now validated against the specified outcome type.
• match.arg() is now used on all categorical arguments.
• Added a check for datasets that are too small to cross-validate (srr G5.8a).

Documentation and vignettes

• New vignette on glmnet and HAL approaches to determining weights.
• Updated the basic examples vignette to be more illustrative, and clarified the multiclass prediction vignette.
• Removed super_learner() from the comparison plot shown by plot.nadir_sl_model() to avoid overstating performance.
• Many documentation improvements: all arguments documented, ORCID iD format updated, removal of links and ::: usage flagged by CRAN, and shorter examples to respect CRAN's runtime limits.

rOpenSci preparation and internal changes

• Added srrstats standards tags throughout the package in preparation for rOpenSci submission.
• Package passes pkgcheck::pkgcheck(); removed unused code and improved docs in response.
• Added lintr configuration and resolved lints (&&/|| in if, seq_len()/seq_along(), removal of commented-out and unused code, line-width limits); lintr also caught a never-called validation function, which is now wired in.
• Ran styler::style_pkg() across the package.
• Expanded the testing suite substantially, including new tests for cross-fitting methods, learner leakage, character formulas, and input ordering.
• Continuous Integration: removed slow ubuntu-devel/oldrel workflows that spent excessive time installing dependencies.

nadir 0.0.1

• Initial CRAN submission.