Changelog
Source:NEWS.md
cpmr (development version)
Breaking changes
-
cpm()now uses a staged specification created byspec(). The oldthresh_method,thresh_level,kfolds, andbias_correctarguments have been replaced byspec,resamples, and staged helper constructors. Unused arguments passed through...now cause an error rather than being ignored. - Removed the deprecated
confoundsalias. Usecovariatesinstead. - The returned
cpmobject now stores workflow state underspec,settings,predictions, andfolds. Code that accessed the oldreal,pred, orparamsfields must be updated.
Enhancements
- Added
spec()together withcpm_selection_cor(),cpm_construction_strength(), andcpm_model_lm()to make screening, summarization, and outcome-model choices explicit. -
cpm()now defaults to 10-fold assessment; explicit leave-one-out and manual assessment plans remain available throughresamples. -
summary.cpm()andtidy.cpm()report pooled RMSE/MAE alongside correlation metrics. Small assessment folds trigger a warning, make fold-wise correlations unavailable, and identify pooled out-of-fold correlations as descriptive; fold-wise summaries report the number of contributing folds. - Unweighted edge standardization now processes only selected edges when constructing training and assessment features, reducing time and memory use for sparse edge selections.
- Fold execution discards intermediate model state after each fit instead of retaining it until all predictions have been assembled.
- Fold-wise edge storage avoids copying the full edge array for every fold.
- Reworked package documentation and pkgdown navigation around the
spec()+cpm()workflow, including a native getting-started vignette.
Maintenance
- Corrected covariate-adjusted p-value screening degrees of freedom and predictions from rank-deficient nuisance regressions.
- Simplified the internal CPM workflow around fold-based execution and aligned
tests/testthatwith the currentR/modules. - Hardened assessment-fold validation and leakage-safe preprocessing so each fold must retain at least 3 complete-case training observations.
cpmr 0.1.1
CRAN release: 2026-03-11
Enhancements
- Renamed internal constant
networksandincludestocorr_typesandinc_edges, respectively, to better reflect their purpose. In addition, the documentation which mentionednetworkshas been updated to use correlation types to better reflect their meaning. - Reorganized
R/cpm.Rsocpm()remains the entry-point function at the top, followed byprint.cpm(), with internal helpers grouped below. - Improved test architecture with explicit internal-helper coverage and complete-case fold invariants.
Maintenance
- Refactored the internal
cpm()workflow into smaller helpers (normalize_inputs,resolve_include_cases,apply_confounds_regression, etc.) while preserving user-facing behavior. - Removed the obsolete internal alias
regress_counfounds(). - Removed broad
Rfastnamespace import in favor of explicitRfast::calls. - Updated
summary.cpm()edge summarization path to avoid mutating local object fields.
cpmr 0.0.8
CRAN release: 2024-03-14
- Added
return_edgesargument to optionally set how to return edges in the output.
cpmr 0.0.5
- Fixed confounds treatment. Now confounds are used in feature selection but not in model fitting.
cpmr 0.0.4
- Ensure sparsity threshold method work as expect.
- Some other improvements in code quality.