Abstract. Parallel coordinate descent algorithms emerge with the growing demand of large-scale optimization. In general, previous algorithms are usually limited by their divergence under high degree of parallelism (DOP), or need data pre-process to avoid divergence. To better exploit parallelism, we propose a coordinate descent based parallel algorithm without needing of data pre-process, termed as Bundle Coordinate Descent Newton (BCDN), and apply it to large-scale ℓ1-regularized logistic regression. BCDN first randomly partitions the feature set into Q nonoverlapping subsets/bundles in a Gauss-Seidel manner, where each bundle contains P features. For each bundle, it finds the descent directions for the P features in parallel, and performs P -dimensional Armijo line search to obtain the stepsize. By theoretical analysis on global convergence, we show that BCDN is guaranteed to converge with a high DOP. Experimental evaluations over five public datasets show that BCDN can better exploit parallelism and outperforms state-of-the-art algorithms in speed, without losing testing accuracy.

No abstract

No abstract