Abstract：Matrix Factorization (MF) based Collaborative Filtering recommenders proves to be highly accurate and scalable, nonetheless, most models of which have the defect of serial training. It is generally accepted that the models can be more scalable if they are able to training in parallel. In order to attain the goals, a new parallel model based on the Regularized Matrix Factorization (RMF) recommenders, namely Parallel RMF (P-RMF), is proposed. P-RMF works by employing the alternative stochastic gradient decent instead of stochastic gradient decent to train the features, whereby the dependence between user features and item features is removed and then the training process can be parallelized and improved. The experiment results show that this new model is more effective to increase both the scalability and computing speed for solving collaborative filtering problems, compared with the existing similar models.
2013, Vol. 35 
