Skip to contents

Changelog

Source: NEWS.md

xplus 1.1.1

  • Seeded functions (xplus(), xplus_elkan_noto(), xplus_naive(), pu_mask(), simulate_pu_data()) now restore the caller’s RNG state on exit instead of clobbering .Random.seed.
  • New pkgdown article “Seeing PU learning in two dimensions” visualizing linear and nonlinear PU problems with glmnet and ranger probability surfaces.

xplus 1.1.0

  • New PU baselines xplus_elkan_noto() and xplus_naive() expressed on the same backend contract as xplus(), enabling honest head-to-head comparisons: Elkan-Noto rescales P-vs-U probabilities by the estimated labeling propensity (mean or median over labeled positives), and the naive baseline treats all unlabeled instances as negatives. Both return xplus_baseline objects with predict support and a $cutoff.
  • pu_xplus() registers xplus as a parsnip model (engine "xplus", classification mode) so fit_xy()/fit(), predict(type = "prob"), update(), and tune::tune_grid()/finetune::tune_sim_anneal() all work against the backend contract. Registration is automatic on load when parsnip is installed and is idempotent. The backend argument selects the learner; backend-constructor arguments (trees, learn_rate, mtry, min_n, tree_depth, sample_size, loss_reduction) accept tune() placeholders with dials parameter mappings and are ignored with a warning when the selected backend does not accept them. Standard yardstick metrics over observed labels give the PU-valid positive-vs-unlabeled score for free.
  • Pluggable fitting backends: xplus() gains a backend argument accepting "glmnet" (default, historical behavior), "ranger", "gam", or a custom xplus_backend() spec with fit/predict functions. alpha and cv_measure configure the glmnet backend only; non-CV backends record "none" in history$cv_measure.
  • New backend constructors xplus_backend_glmnet(), xplus_backend_ranger(), xplus_backend_gam(), xplus_backend_xgboost(), and xplus_backend_lightgbm(); ranger, mgcv, xgboost, and lightgbm move to Suggests. Boosting backends report out-of-fold fit_predict scores from xgb.cv/lgb.cv on the xplus CV folds, the honest analogue of ranger’s out-of-bag predictions.
  • New xplus_backend_parsnip() adapter wrapping any parsnip model spec (e.g. rand_forest() |> set_engine("ranger")) as an xplus() backend: the spec’s mode flips between classification (binary iterative targets) and regression (fractional final targets), case weights map to hardhat::frequency_weights()/importance_weights(), engine seeds are injected for reproducibility, and fit_predict reuses ranger out-of-bag scores when the engine stores them. parsnip, hardhat, and rlang move to Suggests.
  • Backend capability flags supports_fractional and supports_weights on xplus_backend(): fractional-incompatible learners fail with a clear error when the final pseudo-label targets are soft, and weight-ignoring learners warn once under sampling = "bootstrap". Both default to TRUE, preserving the existing contract.
  • New pu_score() and xplus_grid() for learner selection without negative labels: the PU-valid score is the positive-vs-unlabeled AUC of the honest stored predictions, which ranks learners consistently under SCAR. xplus_grid() compares backend names/specs or evaluates a constructor over a params data-frame grid on shared seeds, returning a ranked leaderboard with per-candidate errors captured.
  • simulate_pu_data() gains "xor" (checkerboard sign interactions) and "mixture" (multimodal positive clusters) boundaries, and new pu_mask() converts any open binary dataset into PU labels under SCAR — used by inst/benchmarks/realdata on UCI spambase and mushroom.
  • New get_importance() returning backend feature scores (elastic-net coefficients for glmnet, variable importance for ranger, smooth-term significance for gam). coef() remains glmnet-only.
  • New simulate_pu_data() synthetic PU generator with linear, interaction and nonlinear boundaries under SCAR labeling for benchmark experiments.
  • predict() accepts non-glmnet backends; s stays glmnet-specific. summary()/print() report the backend; predicted_coefficients is a sparse zero column for backends without coefficients.
  • Backends may declare fit_predict(model, x) for honest (out-of-bag or held-out-fold) predictions on their fitted rows; xplus() uses them after the final fit for the stored predictions and classification cutoff, with non-finite entries falling back to predict. The ranger backend implements it via ranger’s out-of-bag predictions, fixing a cutoff calibrated on memorized in-sample scores that no prediction could reach (F1 collapsed to zero on hidden positives). predict() without newx returns the stored honest training predictions on non-glmnet backends.

xplus 1.0.2

CRAN release: 2026-09-26

  • Resubmission to CRAN. Fixed the reviewer notes from Konstanze Lauseker: removed \examples from the unexported internal functions new_xplus() and validate_xplus(); the package no longer uses \dontrun{} anywhere; and xplus() no longer reads or writes .GlobalEnv directly (the optional seed argument uses a plain set.seed() call).

xplus 1.0.1

  • Resubmission to CRAN. Fixed NOTEs reported in the incoming pre-test: reworded Description to avoid false-positive spelling flags and removed the private development repository URL from NEWS.md and README.md.

xplus 1.0.0

This is a major, breaking upgrade of the PLUS-derived package extensions, not an exact reproduction of the paper or evidence of predictive superiority. Response-scale model scores are not guaranteed to be calibrated probabilities. The package remains xplus; the production repository is https://github.com/alrobles/xplus. Development, audit harnesses and historical baselines live in a separate development repository.

Breaking interfaces and metrics

  • Require R >= 4.1.0 for native-pipe syntax and declare glmnet >= 4.1-8. Verification uses R 4.5.2 / glmnet 5.0; other supported versions have not thereby been verified.
  • Reject learning_rate = 0; the accepted interval is (0, 1].
  • Interpret binary factors by labels, not level codes. Single-lambda class predictions now have fixed factor levels "0", "1"; multiple lambdas retain a 0/1 matrix. Classification uses a strict > object$cutoff comparison.
  • Accept positional newx in predict(). Validate finite numeric features; named training features require an exact set of nonempty unique prediction names and are automatically aligned. Reject mismatches and unsupported dots.
  • Respect s for training and new-data prediction, including numeric lambdas. Cache-only objects support only "lambda.min" without new data. get_auc() forwards prediction arguments and requires one lambda; assess() can return per-lambda metrics.
  • Require finite binary truth, or finite nonnegative two-column negative/positive masses for soft truth or counts (assess()/auc_matrix()). Soft vectors are not binary truth. Weights must be finite, nonnegative and row-aligned, without recycling. Invalid rows are errors, not silently dropped observations.
  • Undefined AUC returns NA with a warning when either effective class mass is zero. Zero effective total mass/weight makes assessment undefined. Ties receive half credit; constant scores with both classes have AUC 0.5.
  • MSE and MAE retain the two-class-column loss convention: twice scalar binary loss. Class-mass totals contribute to effective row weights. Deviance clips probabilities to [1e-5, 1 - 1e-5]; other metrics do not inherit that clipping.

Sampling, CV and stopping

  • Default sampling = "bootstrap" preserves replacement multiplicities as case weights on unique rows. Duplicate identities never cross CV folds. sampling = "unique" retains legacy deduplication, not all legacy behavior.
  • sample_use_time is a budget of completed inclusion rounds per identity, not bootstrap draws. Record both sampling_counts and draw_counts.
  • Default cv_measure = "deviance" in both iterative and final fitting stages. AUC remains opt-in, with an explicit warning and deviance fallback for fewer than 10 observations per fold on average; iterative effective measures appear in history.
  • Require at least three positive and three unlabeled input observations. Reduce requested folds to available class counts; require three unique members of each iterative training class and retry degenerate samples at most 20 times before reporting stop_reason = "degenerate_labels".
  • Preserve the selected global-update, hard-thresholded iterative training path when learning_rate < 1; learning_rate = 1 uses sampled Bernoulli updates. Final targets remain soft in both paths. Reference sampling budgets and smoothing should not be mistaken for newly invented package extensions.
  • Expose sigmoid_scale = 10 separately from glmnet elastic-net alpha.
  • Default min_iter = 5, stability_window = 5, min_coverage = 0.9. Stability compares undamped mapped scores against previous pseudo-labels over all unlabeled cases. A full window of consecutive scores strictly above convergence_threshold and minimum coverage are required; damping alone cannot create convergence. Record max_iter, budget_exhausted, degenerate_labels and label_stability distinctly rather than implying every stop is success.

Final fitting and reproducibility

  • Validate effective class mass on the full final dataset and every CV training split (mean positive proportion strictly within 1e-5 and 1 - 1e-5). Soft labels all above 0.5 remain valid when mass is sufficient; hard-class counts do not trigger final fallback.
  • Warn when reverting to original PU labels and record fallback_used and fallback_reason. Preserve original_y, proposed pseudo_labels, actual final_labels/y, iteration history, sampling/draw counts and final_foldid. Insufficient original-label mass is an error, not a hidden recovery.
  • Isolate an explicit seed from the caller’s RNG stream and restore its prior state, including when no .Random.seed existed.

Documentation and verification

  • Regenerate reference manuals from current roxygen contracts. Installation examples target alrobles/xplus without embedded credentials.
  • Diagnostic harnesses, the historical numerical-audit baseline and candidate diagnostics are maintained in a separate development repository; they are not part of the production source tree.
  • Document Rscript -e 'devtools::test(stop_on_failure = TRUE)' and require inspection of FAIL/WARN/SKIP counts. Keep build/check output outside the tracked checkout.

xplus 0.1.1

  • Fixed a crash when the pseudo-labels of the iterative training subset collapse to a single class: fitting now stops cleanly with stop_reason = "degenerate_labels" instead of failing inside glmnet::cv.glmnet().
  • Sampling probabilities are now clamped at zero when the unlabeled-sampling budget is decremented, and sampling is skipped once the budget is fully exhausted (avoids an uninformative sample() error).

xplus 0.1.0