NEWS

---

IntegratedLearner 0.99.6

Survival IBS fusion

• Fixed sample alignment in IBS weight optimization: layer survival matrices now use the same time-ordering permutation as outcomes and censoring weights. Previously, unsorted training data could pair predictions with another sample's outcome and yield incorrect fusion weights.
• Added regression coverage comparing learned weights with an independently evaluated IPCW Brier objective and checking sample-order invariance of weights and fused survival curves for fixed layer predictions, including tied times, censoring, multiple layers, and optional shrinkage toward uniform weights.
• Removed validation-outcome recalibration from survival probability scoring. Single-layer, early-fusion, Cox-fusion, and cooperative models now reuse calibration fitted only to training outcomes and OOF risks; validation predictions preserve the training Cox coefficients, baseline survival, missing-risk replacement, and Brier time grid.
• Compute training and validation IBS-fusion Brier scores directly from the blended survival matrices, without fitting an additional model to their scalar risk summaries. AUC and concordance retain their existing risk-marker definitions and evaluation grids.
• Added brier_time_grid, stored survival_calibration objects, and train_surv/valid_surv matrices for inspecting the probabilities actually evaluated. The metric helper no longer fits calibration implicitly.
• Added regression tests for frozen/serialized calibration, validation-label independence of probabilities, direct IBS scoring, and cooperative survival calibration. Updated the help, README, vignette, and manuscript accordingly. Rerun affected earlier survival probability-error analyses before reporting results; the package version remains 0.99.6 for these corrections.

IntegratedLearner 0.99.5

Multiclass stacking

• Grouped multiclass results under survival-style train_out and valid_out entries for single, early, and all enabled late methods. Each method includes probabilities, classes, and metrics, with final and fold-specific weights for learned averaging. Prediction returns the same validation layout; existing flat outputs are preserved. Added late_methods to enumerate outputs.
• Added majority voting (majority) and an out-of-fold log-loss-weighted probability average (weighted_average) to multiclass fitting, prediction, and metrics. Both default to enabled and can be disabled independently with run_majority and run_weighted. Majority voting returns vote fractions and resolves ties by training class order. Final averaging weights are stored in model_fits$model_weighted$weights; training predictions cross-fit these weights across the existing folds.
• Fit the final multiclass meta-model on out-of-fold layer probabilities, matching the inputs used to train the cross-validated stack and reducing overfitting from in-sample base predictions.
• Updated the README, vignette, and fitting, prediction, and plotting help for all multiclass fusion methods, grouped results, and evaluation caveats.
• Clarified the existing multiclass class-association importance calculation and per-layer grouping. Corrected documentation to describe feature-by-class matrices and an unsigned global ranking; the calculation is unchanged.

IntegratedLearner 0.99.4

Tutorial data and documentation

• Removed the data/ directory, raw tables, duplicate PCL datasets, and their obsolete dataset help pages. Seven MAE-only RDS fixtures in inst/extdata total 257,684 bytes, down from 13,720,424 bytes of packaged datasets.
• Each assay retains at most 25 finite, variable features selected using only training assays in original feature order. Paired validation assays use the same feature IDs and order. All original samples, subject IDs, outcomes, sample maps, and retained assay values are preserved.
• Updated the README, vignette, examples, and tests to access the tutorial RDS files directly with readRDS(system.file("extdata", ...)), replacing the former data() calls without adding a package-specific loader function. PCL inputs remain supported without bundling duplicate PCL datasets.
• Added a deterministic fixture-generation script and an inventory with source provenance and regeneration instructions in inst/extdata/README.md.
• Removed the outdated README project-status paragraph and documented the public entry points and naming conventions. Existing low-level exported names remain available for compatibility; package learner wrappers use sl_*.
• The default tutorial uses non-Java learners, including ranger with multinomial stacking for the small multiclass example. BART and its uncertainty examples remain optional extensions enabled with IntegratedLearner.tutorial_bart.

Dependencies and S3 methods

• Registered update.learner as an S3 method and imported the stats::update generic. Updated its example and tests to call update(fit, ...).
• Declared randomForestSRC in Imports as a required survival backend and replaced its runtime availability guard and dynamic namespace lookup with a direct randomForestSRC::rfsrc() call.
• Removed redundant availability checks for packages declared in Imports and the unused dependency-check helpers.
• Consolidated global-variable declarations in R/globals.R and removed the obsolete R-version check.

Readability and shared implementation

• Documented why make_sl_env() provides package-local learner lookup for SuperLearner without attaching the package or modifying upstream namespaces.
• Removed set_seed_internal() and called withr::local_seed() directly. Centralized and documented the historical cooperative seed offset.
• Consolidated final glmnet fitting across sl_glmnet2(), sl_lasso(), and sl_enet(), preserving the elastic-net alpha search. Shared roxygen sections are inherited; each learner retains its own example.
• Shared cooperative matrix construction, model tuning, and result metadata across outcome families while retaining their distinct responses, folds, filtering, metrics, and output structures. Reused computed feature importance.
• Reused slice_feature_table() for the six cited multiclass transpositions and shared numeric screener-option parsing.
• Consolidated prediction feature checks with exact checks of feature names, lengths, and order to avoid accepting mismatches through vector recycling.
• Removed redundant numeric conversions, multiplication by 1, and no-op null fallbacks. Numeric percentage results may use integer or double storage.
• Removed unused extract_multiclass_screen_args(), map_multiclass_screener(), and ibs_measure() helpers and tests that only exercised deleted helpers.
• Flattened the nesting identified by the reviewer, simplified fragmented comment banners, and clarified section comments without splitting source files.
• Reused MAE preparation in tests and removed the legacy survival-fixture conversion/fallback code and redundant fixture-data loader.

Performance and validation

• Replaced NNLS transpose/matrix products with crossprod().
• Shared and vectorized IPCW Brier calculations. IBS optimization now caches its layer design matrix and censoring weights rather than constructing a three-dimensional array at every evaluation, and uses stable softmax weights.
• Added regression coverage for RNG restoration and reproducible folds, glmnet-wrapper predictions, exact feature alignment, fixture size and structure, vectorized Brier/IBS equivalence, and randomForestSRC fitting and prediction using the required backend.
• Verified retained MAE contents against their source objects. Executed the complete default tutorial across binary, continuous, multiclass, and survival workflows without warnings or errors; optional BART extensions were excluded.
• The revised implementation passes the installed-package test suite (627 assertions, no failures or skips) and the requested unnecessary-nesting linter.

IntegratedLearner 0.99.3

Cooperative multiview learning

• Added run_intermediate as a standalone cooperative multiview learning mode powered by multiview. In this mode, continuous/binary and survival fits use direct feature-level cooperative learning and skip base learners, early fusion, late fusion, and supervised screening.
• Multiclass outcomes do not support cooperative multiview learning; if run_intermediate = TRUE, a message is printed and the option is ignored.
• Added cooperative tuning controls cooperative_rho, cooperative_s, and cooperative_type_measure. rho is selected across repeated cv.multiview() fits with shared folds, while lambda selection is handled by cv.multiview().
• Added coefficient-based cooperative feature-importance outputs, including absolute and signed global scores plus per-layer score lists for continuous/binary and survival cooperative fits.
• Updated the README, vignette, and Rd files to document standalone intermediate/cooperative mode and its output structure.
• Added tests for gaussian, binomial, survival, and multiclass run_intermediate behavior, including checks that cooperative mode does not depend on base or meta learners.

IntegratedLearner 0.99.2

Review-driven updates

• Revised package metadata for the current Bioconductor review cycle.
• Added Himel Mallick's ORCID to the Authors@R field in DESCRIPTION.
• Replaced the previous unformatted news file with a versioned NEWS.md.

Bioconductor workflow and documentation

• Reworked the package documentation and vignette to make MultiAssayExperiment the primary user-facing workflow.
• Expanded the vignette introduction to describe how IntegratedLearner fits into a typical Bioconductor pipeline.
• Updated examples and vignette code to use standard Bioconductor containers and accessors such as MultiAssayExperiment, SummarizedExperiment, experiments(), assay(), colData(), and sampleMap().
• Added package-level documentation and clarified return values for exported methods, including predict.learner().
• Revised man page examples to be runnable under package checking and removed blank if (FALSE) example blocks.
• Removed disabled vignette install chunks and switched installation snippets to plain code blocks to reduce the risk of documentation drift.

Dependencies and installation

• Reorganized dependencies so that packages needed at runtime are listed in Imports, while vignette/example/test-only packages are kept in Suggests where appropriate.
• Removed the runtime package-code dependence on stringr and replaced those uses with base R equivalents.
• Updated README installation instructions to include Bioconductor usage.

API and naming cleanup

• Continued aligning package-specific exported helper/learner names with snake_case conventions and reserving . for S3 methods.
• Preserved backward compatibility for older learner identifiers while supporting the newer naming style.

Examples, testing, and package checks

• Updated exported examples for IntegratedLearner, predict.learner(), update.learner(), and plot.learner() to improve reproducibility under R CMD check.
• Removed rendered vignette artifacts from vignettes/ and added ignore rules to avoid recommitting generated HTML outputs.
• Expanded and refreshed tests around model methods, MAE workflows, prediction/update behavior, and utility helpers.

Internal code quality

• Replaced the previous internal seed handling approach with a cleaner withr-based implementation.
• Updated print.learner() to return invisibly, consistent with base print() conventions.
• Reduced code repetition in the continuous/binary training, prediction, and update paths by introducing shared helper utilities for validation-layer prediction assembly, fusion prediction assembly, and family-specific metric calculation.