An efficient augmented Lagrangian-based method for linear equality-constrained Lasso

Abstract

Variable selection is one of the most important tasks in statistics and machine learning. To incorporate more prior information about the regression coefficients, various constrained Lasso models have been proposed in the literature. Compared with the classic (unconstrained) Lasso model, the algorithmic aspects of constrained Lasso models are much less explored. In this paper, we demonstrate how the recently developed semis-mooth Newton-based augmented Lagrangian framework can be extended to solve a linear equality-constrained Lasso model. A key technical challenge that is not present in prior works is the lack of strong convexity in our dual problem, which we overcome by adopting a regularization strategy. We show that under mild assumptions, our proposed method will converge superlinearly. Moreover, extensive numerical experiments on both synthetic and real-world data show that our method can be substantially faster than existing first-order methods while achieving a better solution accuracy.