[R-pkgs] sparsenet: a new package for sparse model selection

[Trevor Hastie]

We have put a new package sparsenet on CRAN.

Sparsenet fits regularization paths for sparse model selection via coordinate descent, 
using a penalized least-squares framework and a non-convex penalty.

The package is based on our JASA paper
Rahul Mazumder, Jerome Friedman and Trevor Hastie: SparseNet : Coordinate Descent with Non-Convex Penalties. (JASA 2011)
http://www.stanford.edu/~hastie/Papers/Sparsenet/jasa_MFH_final.pdf

We use Zhang's  MC+ penalty to impose sparsity in model selection. This penalty 
parametrizes a family ranging between L1 and L0 regularization. One nice feature of this
family is that the single-coordinate optimization problems are convex, making it
ideal for coordinate descent.

The package fits the regularization surface for each parameter -  a surface over the 
two-dimensional space of tuning parameters. The concavity parameter gamma indexes 
the member of the family, and lambda is the usual Lagrange penalty parameter which
determines the strength of the penalty.

Sparsenet is extremely fast. For example, with 10K variables and 1K samples, the entire surface with
10 values of gamma and 50 values of lambda takes under a second on a Macbook Pro.

The package includes functions for fitting, plotting and cross-validation of the models,
as well as methods for prediction.

Trevor Hastie, with Jerome Friedman and Rahul Mazumder