Sparse Least-Squares Methods in the Parallel Machine Learning (PML) Framework

We describe parallel methods for solving large-scale, high-dimensional, sparse least-squares problems that arise in machine learning applications such as document classification. The basic idea is to solve a two-class response problem using a fast regression technique based on minimizing a loss function, which consists of an empirical squared-error term, and one or more regularization terms. We consider the use of Lenclos-based methods for solving these regularized least-squares problems, with the parallel implementation in the Parallel MachineLearning (PML) framework, and performance results on the IBM Blue Gene/P parallel computer.