NEWS

---

PsyML 2.4.2.9000

• Stratified splits now keep rare classes stratified: rsample pools strata that hold less than 10% of the data, so with a rare outcome the train/test split and the cross-validation folds were not stratified and the number of positive cases varied between partitions. Classification splits now use pool = 0.
• balance_method = "down" failed in every model ("All models failed") because themis::step_downsample() was called with over_ratio instead of under_ratio.

PsyML 2.4.2 (2026-11-12)

First release under the name PsyML, prepared for CRAN.

• The package was renamed from easyML (version 2.4.2) to PsyML.
• The package is organized around three approaches, each run by one function: supervised learning (supervised_ml()), unsupervised learning (unsupervised_ml()) and causal inference (causal_ml()). The package help page (?PsyML), the new vignette "Three Approaches" and the "Other supervised learning", "Other unsupervised learning" and "Other causal inference" links in each help page show which functions belong to each approach.
• The reinforcement learning module (reinforcement_ml() and its methods, bandits, MDP and tabular agents, which depended on the contextual package) was removed.
• Variable importance no longer uses the archived vip package: get_importance() now uses model-specific measures (impurity for ranger and rpart, gain for xgboost, absolute coefficients for glmnet, absolute z/t statistics for glm/lm) and falls back to permutation importance implemented in the package.
• SHAP values are computed with kernelshap (Suggests) instead of the archived fastshap package; shapviz (Suggests) is still used for plots when installed.
• All documentation, the vignette, the README and the condition messages (errors, warnings, messages) are now in English. Console progress output of the pipelines is still in Spanish.
• seed arguments now default to NULL: the random number generator is only seeded when a seed is supplied. supervised_ml() uses 2 cores by default (n_cores = 2).
• Nothing is written to disk by default: export_verbose_txt(), export_verbose_json(), generate_report_from_json(), supervised_ml_export(), save_all_plots(), save_unsupervised_plots() and generate_report_with_ai() require an explicit path.
• explain_with_ai() and generate_report_with_ai() no longer include a shared API key; they read the key from api_key or from the environment variable PSYML_API_KEY. explain_with_ai() gains a verbose argument.
• predict.unsupervisedml() now takes object as its first argument, and Gaussian mixture clustering no longer attaches mclust.
• cluster_profile() gains a verbose argument.
• Imports were reduced to the packages actually used (the tidyverse and tidymodels meta-packages are no longer imported).

• fit_models_cv(): in binary classification the cross-validated PR-AUC was computed for the negative class, because tune::fit_resamples() passes the first probability column to the metrics and pr_auc did not follow the event_level. With imbalanced classes this gave implausible values (e.g. .999 with 4% positives versus a true .786). It is now recomputed per fold with the probability of the positive class. ROC-AUC was not affected.

PsyML 2.4.1

Fixes found while running supervised_ml() on data with strongly imbalanced classes (4% positives) and highly collinear predictors.

• compare_models(): supervised_ml() with a single model failed when computing the average ranking.
• calculate_shap(): predictors removed by preprocessing (high correlation, VIF, zero variance) appeared in the SHAP importance with value 0; they are now excluded and reported in $removed_vars.
• calculate_shap(): new $collinear_vars with the predictors used that have VIF > 5, and a note that the split of SHAP values among nearly redundant predictors should not be interpreted variable by variable.
• calculate_shap(): in classification the SHAP sample is stratified with the same number of cases per class ($sample_design).
• Optimal threshold: it is now chosen with out-of-fold predictions on the training set and evaluated on the test set ($threshold_optimization$source, $threshold_optimization$test_metrics). Previously it was chosen and evaluated on the same test predictions.