# BeSS 2.0.0

## Improvements

• The majority of code in the package is now written in C++ for fast implementation, in which linear algebra is supported by the Eigen3 library for portable, high-performance computation. With Rcpp and RcppEigen, the C++ program can be called from R by user-friendly interfaces.

## New features

• It supports the best-subset ridge regression model (BSRR), which is a more flexible model. The advantage of BSRR is that it can adapt to the low signal-to-noise ratio and high-multicollinearity setting, tackle the non-identifiable problem in BSS when $$p>n$$, and often outperforms the best subset selection and other variable selection algorithms in terms of prediction accuracy while maintaining the model’s parsimony at the same time; To realize a BSRR model, set the type bsrr.

• The Poisson model is added as a complement for the case where the response is a count value.

• The BeSS package now can select variables with group structures (e.g., a group of dummy variables corresponding to a multilevel variable). The option group.index indicates the group index for each variable.

• Sure independent screening can be carried out through the option screening.num in consideration of dealing with the ultra-high dimensional data.

• It is now possible to force specified variables in the model through the option always.include which takes the index of the variables that should always be included.