NEWS

abclass 0.4.0 (2022-09-18)

Added support of sparse matrix x of class sparseMatrix (provided by the {Matrix} package) for abclass() and predict.abclass().
Added new functions named cv.abclass() and et.abclass() for training and tuning the angle-based classifiers with cross-validation and an efficient tuning procedure for lasso-type algorithms, respectively. See the corresponding function documentation for details.
Added experimental classifiers with sup-norm penalties. See the functions supclass() and cv.supclass() for details.

Simplified the function abclass() and moved the tuning procedure by cross-validation to the function cv.abclass().

Changed the default values of the following arguments for abclass.control().
- alpha: from 0.5 to 1.0
- epsilon: from 1e-3 to 1e-4

Added experimental group-wise regularization by group SCAD and group MCP penalty.
Added a new function named abclass.control() to specify the control parameters and simplify the main function interface.

Removed the function call from the return of abclass() to avoid unnecessarily large returned objects
Changed the default value of lum_c for abclass() from 0 to 1.
Renamed the argument rel_tol to epsilon for abclass().

Fixed the first derivatives of the boosting loss
Fixed the label prediction by using the fitted inner products instead of the probability estimates
Fixed the computation of regularization terms for verbose outputs in AbclassNet
Fixed the computation of validation accuracy in cross-validation
Fixed the assignment of lum_c in the associated header files.
Fixed the computation of lower bound for distinct observation weights

The first release of abclass providing the multi-category angle-based large-margin classifiers with various loss functions.