ADMM Algorithmic Regularization Paths for Sparse Statistical Machine Learning
Optimization approaches based on operator splitting are becoming popular for solving sparsity regularized statistical machine learning models. While many have proposed fast algorithms to solve these problems for a single regularization parameter, conspicuously less attention has been given to computing regularization paths, or solving the optimization problems over the full range of regularization parameters to obtain a sequence of sparse models. In this chapter, we aim to quickly approximate the sequence of sparse models associated with regularization paths for the purposes of statistical model selection by using the building blocks from a classical operator splitting method, the Alternating Direction Method of Multipliers (ADMM). We begin by proposing an ADMM algorithm that uses warm-starts to quickly compute the regularization path. Then, by employing approximations along this warm-starting ADMM algorithm, we propose a novel concept that we term the ADMM Algorithmic Regularization Path. Our method can quickly outline the sequence of sparse models associated with the regularization path in computational time that is often less than that of using the ADMM algorithm to solve the problem at a single regularization parameter. We demonstrate the applicability and substantial computational savings of our approach through three popular examples, sparse linear regression, reduced-rank multi-task learning, and convex clustering.