English

A Risk Ratio Comparison of $l_0$ and $l_1$ Penalized Regression

Statistics Theory 2015-10-22 v1 Methodology Statistics Theory

Abstract

There has been an explosion of interest in using l1l_1-regularization in place of l0l_0-regularization for feature selection. We present theoretical results showing that while l1l_1-penalized linear regression never outperforms l0l_0-regularization by more than a constant factor, in some cases using an l1l_1 penalty is infinitely worse than using an l0l_0 penalty. We also show that the "optimal" l1l_1 solutions are often inferior to l0l_0 solutions found using stepwise regression. We also compare algorithms for solving these two problems and show that although solutions can be found efficiently for the l1l_1 problem, the "optimal" l1l_1 solutions are often inferior to l0l_0 solutions found using greedy classic stepwise regression. Furthermore, we show that solutions obtained by solving the convex l1l_1 problem can be improved by selecting the best of the l1l_1 models (for different regularization penalties) by using an l0l_0 criterion. In other words, an approximate solution to the right problem can be better than the exact solution to the wrong problem.

Keywords

Cite

@article{arxiv.1510.06319,
  title  = {A Risk Ratio Comparison of $l_0$ and $l_1$ Penalized Regression},
  author = {Kory D. Johnson and Dongyu Lin and Lyle H. Ungar and Dean P. Foster and Robert A. Stine},
  journal= {arXiv preprint arXiv:1510.06319},
  year   = {2015}
}