A Note on Application of Nesterov’s Method in Solving Lasso-Type Problems

Ye Tian,H. Lian
DOI: https://doi.org/10.1080/03610918.2013.824096
2015-04-09
Abstract:Many different algorithms have been proposed to solve penalized variable selection problems, in particular lasso and its variants, including group lasso and fused lasso. Loss functions other than quadratic loss also pose significant challenges for finding efficient solvers. Here, we note that Nesterov’s method can be used to transform an optimization problem with general smooth convex loss to quadratic loss with identity covariate matrix in each iteration. After such reduction, the problem becomes much easier to solve or even can be solved in closed form in some cases. We perform some simulations and apply our implementation to phoneme discrimination.
What problem does this paper attempt to address?