[R] glmnet_1.5 uploaded to CRAN
hastie at stanford.edu
Thu Nov 4 20:42:52 CET 2010
This is a new version of glmnet, that incorporates some bug fixes and
* a new convergence criterion which which offers 10x or more speedups for
saturated fits (mainly effects logistic, Poisson and Cox)
* one can now predict directly from a cv.object - see the help files for cv.glmnet
* other new methods are deviance() for "glmnet" and coef() for "cv.glmnet"
Here is the description of the package.
glmnet is a package that fits the regularization path for linear, two- and multi-class logistic regression
models, poisson regression and the Cox model, with "elastic net" regularization (tunable mixture of L1 and L2 penalties).
glmnet uses pathwise coordinate descent, and is very fast.
Some of the features of glmnet:
* by default it computes the path at 100 uniformly spaced (on the log scale) values of the regularization parameter
* glmnet is very fast, even for large data sets.
* recognizes and exploits sparse input matrices (ala Matrix package). Coefficient matrices are output in sparse matrix representation.
* penalty is (1-a)*||\beta||_2^2 +a*||beta||_1 where a is between 0 and 1; a=0 is the Lasso penalty, a=1 is the ridge penalty.
For many correlated predictors, a=.95 or thereabouts improves the performance of the lasso.
* convenient predict, plot, print, and coef methods
* variable-wise penalty modulation allows each variable to be penalized by a scalable amount; if zero that variable always enters
* glmnet uses a symmetric parametrization for multinomial, with constraints enforced by the penalization.
* a comprehensive set of cross-validation routines are provided for all models and several error measures
* offsets and weights can be provided for all models
Examples of glmnet speed trials:
Newsgroup data: N=11,000, p= 0.75 Million, two class logistic. 100 values along lasso path. Time = 2mins
14 Class cancer data: N=144, p=16K, 14 class multinomial, 100 values along lasso path. Time = 30secs
Authors: Jerome Friedman, Trevor Hastie, Rob Tibshirani.
See our paper http://www-stat.stanford.edu/~hastie/Papers/glmnet.pdf for implementation details,
and comparisons with other related software.
Trevor Hastie hastie at stanford.edu
Professor, Department of Statistics, Stanford University
Phone: (650) 725-2231 (Statistics) Fax: (650) 725-8977
(650) 498-5233 (Biostatistics) Fax: (650) 725-6951
address: room 104, Department of Statistics, Sequoia Hall
390 Serra Mall, Stanford University, CA 94305-4065
More information about the R-help