Original author and previous maintainer, Martin Vincent, has stopped development. This version is released by a new maintainer. Changes from version 2.3.6 are only to comply to CRAN policies.
API changes: msgl.lambda.seq deprecated use msgl::lambda, msgl deprecated use msgl::fit, msgl.cv deprecated use msgl::cv, msgl.subsampling deprecated use msgl::subsampling
fit, cv and subsampling supports relative lambda
Code cleanup, bug fixes and refactoring
Removed non-generic default sample weights from cv – this may imply that the lambda sequence needs adjustment to reach the same results.
added quick-start vignette
added functions for extracting number of nonzero features and parameters (
seed argument removed
minor fixes and updates
Better handling of R memory protection (fewer PROTECT calls)
subsampling uses foreach and doParallel (showing a progressbar is no longer supported for subsampling and cross validation)
subsampling now supports different lambda sequences for each subsample
Improved reliability and usability: prefix printing; correct handling of non penalized parameters when computing lambda.max (fit non penalized parameters before computing critical lambdas); increased speed when optimizing with a high number of groups/classes (> 100); some improvements on error reporting and argument validation
fixed bug related to grouping
fixed seval minor bugs
fixed startup messages