Empirical Study of Matrix Factorization Methods for Collaborative Filtering

Matrix factorization methods have proved to be very efficient in collaborative filtering tasks. Regularized empirical risk minimization with squared error loss function and L2-regularization and optimization performed via stochastic gradient descent (SGD) is one of the most widely used approaches. The aim of the paper is to experimentally compare some modifications of this approach. Namely, we compare Huber's, smooth epsilon-insensitive and squared error loss functions. Moreover, we investigate a possibility to improve the results by applying a more sophisticated optimization technique - stochastic meta-descent (SMD) instead of SGD.