GWMF: Gradient weighted matrix factorisation for recommender systems

Abstract

In this paper, we developed a new algorithm, Gradient Weighted Matrix Factorisation (GWMF), for matrix factorisation. GWMF uses weights to focus the approximation in the matrix factorisation to the higher approxmation residual. Therefore, it improves the matrix factorisation accuracy and increases the speed of convergence. We also introduce a regularisation parameter to control overfitting. We applied our algorithm to a movie recommendation problem and GWMF performs better than ordinal gradient descent-based matrix factorisation (GMF) on Movielens dataset. GWMF converges faster than GMF and it guarantees lower root mean square error (RMSE) at earlier iterations on both training and testing. © 2013 Springer-Verlag.