A greedy homotopy method for regression with nonconvex constraints
Wauthier FL., Donnelly P.
Copyright 2015 by the authors. The goal of this paper is to estimate sparse linear regression models, where for a given partition G of input variables, the selected variables are chosen from a diverse set of groups in G. We consider a novel class of nonconvex constraint functions, and develop RepLasso, a greedy homotopy method that exploits geometrical properties of the constraint functions to build a sequence of suitably adapted convex surrogate problems. We prove that in some situations RepLasso recovers the global minima path of the noncon-vex problem. Moreover, even if it does not recover the global minima, we prove that it will often do no worse than the Lasso in terms of (signed) support recovery, while in practice outperforming it. We show empirically that the strategy can also be used to improve over various other Lasso-style algorithms. Finally, a GWAS of ankylosing spondylitis highlights our method's practical utility.