Changes since 1.00 2022/04/08: 1.10-1 . Added citations to DESCRIPTION and to help file for randomGLM 2021/06/12: 1.10 . Bump version for release 2021/05/30: 1.09-992 . Minor changes in internal code handling multi-threading (via cluster functions from parallel) on Windows. . Cleanup of metadata (DESCRIPTION) and help files to comply with CRAN rules . New argument 'weight' lets the user specify sample weights for the underlying regression models. . DIFF: randomGLM now works with categorical data. Categorical variables are first binarized, then used as other numeric variables except their power in interactions can never be more than 1. This may lead to slightly different predictions when some features are categorical. . New arguments categoricalColumns and maxCategoricalLevels allow the user to specify which columns should be treated as categorical. . New argument includeSelfinteractions allows the user to turn off self-interactions for continuous features. Older versions of randomGLM allowed self-interactions but other statistical approaches usually exclude them. . New argument replaceBadBagFeatures optionally turns on a more efficient handling of features with missing data. The old behaviour was to remove them from each bag without attempting to replace them; the new, optional behaviour is to only sample from features that have no missing values in the bag samples. This should lead to more in-bag variables when missing data are present. . Bug fix: randomGLM now correctly records variables that were retained by the model selection when missing data are present. This bug led to incorrect results from functions predict and thinRandomGLM. . Function accuracyMeasures now also works with factors. 2013/05/22: 1.05-2 . Bugfix: function works correctly with suvival objects. . Internal bugfix that may lead to slightly different sets of features being selected for each bag. 2013/05/09: 1.05-1 . More descriptive error message generated if all features are removed from a bag due to missing data or constantness. . Bug fix: predictor now works with a test set that consists of a single sample. 2013/02/18: 1.05 . The predictor now works with a variety of responses including all glm families and censored time (Surv) responses. 2013/02/06: 1.03 . New function accuracyMeasures for convenient calculation of most accuracy measures. 2013/01/28: 1.02-1 . Fixes to problems with feature names that are not valid R names. . Fixed a problem when using interactions with large numbers of features. This required a switch from the gtools' function combinations to an internal version; hence gtools is not required anymore. 2013/01/23: 1.02 . Coefficient names in returned models and in featuresInForwardRegression are now easier to read for humans. New argument interactionSeparatorForCoefNames lets the user specify how interaction names are formed. 2013/01/20: 1.01 . New argument minInBagObs lets user specify the minimum number of unique in-bag observations for each bag. This prevents bags with too few observations that can cause problems when fitting models. . Function now works even when there is a single OOB observation in some bags 2013/01/15: 1.00-2 . Fixed a failure of multi-threaded calculation with more than 100 bags. 2013/01/13: 1.00-1 . Fixed a bug with default (NULL) nThreads. Changes to version included in WGCNA . Parallelized main loop; number of threads can be controled via argument nThreads . bagObsIndex element of output is now nObservationsPerBag times nBags (transposed to what it was before)