Improving the efficiency of IRWLS SVMs using parallel Cholesky factorization

This paper proposes a new and efficient parallel schema of the Iterative Re-Weighted Least Squares (IRWLS) procedure to solve Support Vector Machines (SVMs). This procedure makes use of a parallel Cholesky decomposition to solve in every iteration the linear systems. In particular, we provide two different solutions, a parallel implementation of the IRWLS procedure (PIRWLS) to solve a full SVM and a new parallel implementation of a semi-parametric model of SVM (PSIRWLS). Both solutions have been implemented for multicore and multiprocessor environments with shared memory. We have benchmarked these algorithms against LibSVM, SVMLight and PS-SVM. Experimental results show that using large datasets, our systems offer better parallelization capabilities and higher speed.