An Efficient Approach of GPU-accelerated Stochastic Gradient Descent Method for Matrix Factorization

Abstract

Matrix Factorization (MF) is a very effective tool for Collaborative Filtering (CF) in recommender systems. As a popular solver, Stochastic Gradient Descent (SGD) is widely utilized to find MF solutions for CF. However, SGD solver often suffers from a very slow optimization process, due to its large computation burden. How to speed up it becomes a very important research topic. One of the main techniques to the problem is partitioning the matrix to factorize into blocks and calculate the factorization parallelly with these blocks. In this paper, we would like to use the most modern computation resource Graphics Processing Unit (GPU) to speed up the partition-based computation. Though there are some studies on partition-based SGD with GPUs, due to the sparsity of matrices in real-life scenarios, these methods produce too many blank blocks, which will waste the GPU computing resources. In this paper, we propose a new method, which can avoid the problem and make use of GPUs more efficiently to speed up the SGD based MF solver.