Efficient parallel Stochastic Gradient Descent for matrix factorization using GPU

Abstract

Matrix factorization is an advanced and efficient technique for recommender systems. Recently, Stochastic Gradient Descent (SGD) method is considered to be one of the most popular techniques for matrix factorization. SGD is a sequential algorithm, which is difficult to be parallelized for large-scale problems. Nowadays, researches focus on efficiently parallelizing SGD. In this research, we propose an efficient parallel SGD method, ESGD, for GPU. ESGD is more efficient than recent parallel methods because it utilizes GPU, reducing non-coalesced access of global memory and achieving load balance of threads. In addition, ESGD does not require any sorting and/or data shuffling as preprocessing phase. Although platform used for ESGD implementation is old, ESGD demonstrates 12.5× speedup over state-of-the-art GPU method, BSGD.