Parallel Coordinate Descent for L1-Regularized Loss Minimization
Machine Learning
2011-05-27 v1 Information Theory
math.IT
Abstract
We propose Shotgun, a parallel coordinate descent algorithm for minimizing L1-regularized losses. Though coordinate descent seems inherently sequential, we prove convergence bounds for Shotgun which predict linear speedups, up to a problem-dependent limit. We present a comprehensive empirical study of Shotgun for Lasso and sparse logistic regression. Our theoretical predictions on the potential for parallelism closely match behavior on real data. Shotgun outperforms other published solvers on a range of large problems, proving to be one of the most scalable algorithms for L1.
Cite
@article{arxiv.1105.5379,
title = {Parallel Coordinate Descent for L1-Regularized Loss Minimization},
author = {Joseph K. Bradley and Aapo Kyrola and Danny Bickson and Carlos Guestrin},
journal= {arXiv preprint arXiv:1105.5379},
year = {2011}
}