Using Stochastic Gradient Descent On Parallel Recommender System with Stream Data

Abstract

Stochastic gradient descent (SGD) and Alternating least squares (ALS) are two popular algorithms applied on matrix factorization. Moreover recent researches pay attention to how to parallelize them on daily increading data. About large-scale datasets issue, however, SGD still suffers with low convergence by depending on the parameters. While ALS is not scalable due to the cubic complexity with the target time rank. The remaining issue, how to operate system, almost parallel algorithms conduct matrix factorization on a batch of training data while the system data is real-time. In this work, the authors proposed FSGD algorithm overcomes drawbacks in large-scale issue base on coordinate descent, a novel optimization approach. According to that, algorithm updates rank-one factors one by one to get faster and more stable convergence than SGD and ALS. In addition, FSGD is feasible to paralleize and operates on a stream of incoming data. The experimental results show that FSGD performs much better in solving the matrix factorization issue compared to existing state-of-the-art parallel models.