Matlab software implementing the algorithms described in these papers:
[1] W. Shi, G. Wahba, S. J. Wright, K. Lee, R. Klein, and B. Klein, "LASSO-Patternsearch Algorithm with Application to Ophthalmology and Genomic Data," Statistics and its Interface 1 (2008), pp. 137-153.
[2] S. J. Wright, "Accelerated Block-Coordinate Relaxation for Regularized Optimization," Technical Report, August 2010. Revised September 2011.
The code does l1-regularized linear logistic regression with on data with Bernoulli outcomes (indicated by +/-1). The algorithm uses a gradient projection / iterative shrinkage approach, with gradient sampling and a modified reduced Newton scaling technique on the space of nonzero variables.
The codes were written initially by W. Shi and S. Wright in 2006-2008 and rewritten for distribution in 2008-2011 by S. Wright.
Updated 9/1/2011
Code and Small Test Data Sets
These zip and tarballs contain Matlab code in subdirectory "code" and test data sets in subdirectory "data." See README.txt in the main directory for a short description of contents.
- LATEST: LPS-v2.2.zip or LPS-v2.2.tar.gz. (Posted 9/1/11.) Issued in conjunction with the revision of [2]. Main change is an "adjustment" strategy in which a reduced Newton method is applied to the solution at the previous value of regularization parameter to obtain a warm start for the current value of tau.
-
LPS-v2.1.zip or LPS-v2.1.tar.gz. (Posted 1/17/11.) Implements the algorithm of [2] which modified the approach of [1] in several ways to accommodate a convergence analysis while retaining performance characteristics of the earlier implementation. For further information, see the README.txt file in the distribution and [2].
- LPS-v1.2.zip or LPS-v1.2.tar.gz. (Posted 5/8/2010.) Implements essentially the algorithm described in [1], with modifications such as Hessian approximations.
Larger Test Data Sets
Unzip, and edit the calling programs (TestBig or TestTables in LPS v2.1) to point to the location of the file on your system.
Updates
See the log of changes for notes on the various releases.
Support
This research was supported by National Science Foundation Grants DMS-0505636, DMS-0604572, SCI-0330538, DMS-0427689, CCF-0430504, DMS-0914524, and DMS-0906818; NIH Grant EY09946; and DOE Grant DE-FG02-04ER25627. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.