NEWS

vennabers 1.0.0 (2026-11-07)

Initial CRAN release.

venn_abers() fits an Inductive Venn-Abers predictor to calibration probabilities and labels from any binary model; cross_venn_abers() combines the predictors of matched cross-fitted models. predict() returns the lower and upper Venn-Abers probabilities, a regularized probability chosen by log loss or Brier score, and class decisions. Calibration probabilities can be rounded with precision.
brier_score(), log_loss() and calibration_plot() assess held-out predictions.

venn_abers() and cross_venn_abers() accept factor truth and named probability matrices, with one-vs-one coupling (default) or one-vs-rest normalization. Class diagnostics are reconstructible from the stored binary subproblems.
venn_abers_one_hot() represents existing class decisions as a 0/1 matrix.

venn_abers_regression() and cross_venn_abers_regression() calibrate real-valued predictions into a point with lower and upper bounds in target units, with moderation set by m or epsilon.

With the optional 'tidymodels' packages installed, calibrated_workflow() trains a model and its Venn-Abers calibrator together: Inductive or Cross, binary, shared one-vs-one or one-vs-rest, pairwise one-vs-one and regression. Calibration splits are created locally (calibration_prop, or independent training_size and calibration_size), or given explicitly through local_calibration_splits(). prediction_model = "full" refits the model on all local rows while keeping the held-out calibrators.
calibrate_workflow() adds a calibrator to an already fitted workflow without refitting it.
model_adapter() turns your own fit and predict functions into a model for workflows. venn_abers_schema(), venn_abers_encode() and venn_abers_decode() keep original column names and class order. venn_abers_details() returns bounds, fold identities and source-model diagnostics.
calibrated_workflow() specifications can be assessed by native outer resampling with tune::fit_resamples() and tune::last_fit().

calibrated_workflow(tuning = ...) selects model, recipe and calibration parameters by Grid or Bayes search, jointly or in two stages, inside each local fit. venn_abers_tuning() returns the stored search diagnostics.

calibrated_workflow(postprocessing = ...) learns a postprocessor on rows of its own before the Venn-Abers calibrator: Sigmoid (Platt) scaling, Tailor logistic, multinomial or linear calibration, or numeric isotonic regression, as a single holdout or an inner ensemble.

plot() methods draw the fitted calibration map or the Venn-Abers width of manual calibrators and workflow fits.
corp_plot() draws CORP reliability diagrams with resampling bands, corp_decomposition_plot() splits the Brier score into miscalibration, discrimination and uncertainty, and regression_plot() shows held-out regression predictions. All graphics use base R.

Guides on the method and its guarantees, manual calibration, workflows, tuning, diagnostics and a comparison with Platt and isotonic calibration, plus four collections of the installed examples (browseVignettes("vennabers")).
citation("vennabers") lists the package and the method papers.

The user-facing functions listed in help(package = "vennabers"), their arguments and defaults, and the output formats their help pages document stay compatible throughout 1.x. Incompatible changes wait for 2.0.0; new options keep the previous behavior unless enabled explicitly. Documented bug fixes may still change numerical results, and NEWS names each one.
The parsnip engine entry points fit_vennabers(), fit_vennabers_proxy() and predict_vennabers_*() are exported only for model registration; use calibrated_workflow() instead.
The minimum R version is not part of this promise. Each release supports at least the previous R minor series and may raise the minimum, noted here.

Versions 0.1.0 to 0.9.0 were internal development stages and were never published. Their change log is kept in tools/NEWS-development.md in the source repository.