Genetic algorithm based parallel matrix factorization for recommender systems

Abstract

Matrix Factorization is one of the popular approaches for learning the latent characteristics from the sparse utility matrix of recommendation systems. In recent times, Coordinate Descent based matrix factorization approach (CCD) have outperformed the other existing approaches such as Alternating Least Squares (ALS) and Stochastic Gradient Descent (SGD). While ALS is not scalable due to its cubic time complexity, SGD suffers from slow convergence. An improved version of CCD, CCD++ was recently proposed to overcome the shortcomings of CCD. The difference in these two approaches lies in their update rules and the update sequences. CCD++ was shown to converge faster than CCD. In this paper, we hypothesize that use of Genetic Algorithm (GA) for initializing matrices may significantly speed up the convergence of CCD++. Also, parallelism could be exploited more efficiently at update stage. We update the rating matrix at regular intervals with GA, so that the convergence of CCD++ is relatively fast. Our experimental results show that optimum update of matrices enhances the convergence of CCD++ appreciably. © 2016 IEEE.