Department of Mathematics
Title : How to regularize with a lasso in a positive manner
| Speaker: Assoc. Prof. David Bryant Affiliation: University of Otago Time: 2 pm Thursday, 18 April, 2013 Location: Room 561-303 |
Abstract
| The Lasso (also known as L1 regularization) is a method which has become especially popular in inverse problem theory and statistics because of its ability to automatically select subsets of variables. The LARS-Lasso algorithm of Efron et al. (2004) makes the lasso computational feasible: with little added computational cost one can construct lasso solutions for all ranges of the shrinkage parameter. In many applications, however, model considerations impose a non-negativity constraint on the variable coefficients, giving a demand for a "positive lasso". Efron et al (2004) propose a LARS type algorithm for the positive lasso, but it is not guaranteed to produce optimal solutions. Here we describe a modification to the LARS algorithm which does produce optimal positive lasso solutions, and illustrate the algorithm using an example from evolutionary biology. |
-
Programmes and Centres
- New Zealand Institute of Mathematics and its Applications (NZIMA)
- Community for Understanding and Learning in the Mathematical Sciences (CULMS)
- Centre for Mathematical Social Science (CMSS)
- Department of Computer Science
- Department of Engineering Science
- Department of Physics
- Department of Statistics
- Auckland Bioengineering Institute
