Study On Parallel Collaborative Filtering Recommender System

Electronic Business is defined as the commercial activities by electronic means.Electronic Business System (EBS) builds virtual on-line stores via Internet. However,with the quick expansion scope of the EBS, customers have to glance over a mass ofuseless information before finally coming across the needed commodities. To solve theinformation overload problem referred to, so as to guide consumers shopping effectivelyand conveniently, recommender system technology is proposed. In the past researches,several recommender algorithms have been put forward to ensure efficient and accuraterecommender results, which have been divided into the following categories:association rule-based recommendation, content-based recommendation, collaborativefiltering recommendation and hybrid recommendation. Among the first threerecommender algorithms, collaborative filtering recommendation is proved to be a veryprosperous algorithm for the great efficiency.In recent years, recommender system is facing the double challenge of vastinformation and real-time recommending, so that this paper aims at proposing a newrecommender model based on technology of CFR, which both has the advantage ofCFR and the parallel computing. The main tasks of the paper are list below:①Studying the main kinds of recommender algorithms in electronic business andanalyzing the advantages and disadvantages of each algorithm, then stating thecollaborative filtering recommender based on Regularized Matrix Factorization (RMF)with the expanding models in detail.②Carrying out the theoretical analysis on the parameter updating process of RMFmodel, whereby figuring out that the main obstacle preventing the model fromparallelism is the inter-dependence between item and user features.③Discussing the Stochastic Gradient Descent (SGD) and Parallel AlternatingLeast Square (P-ALS) in the computing process of the RMF model, and applying theAlternating Stochastic Gradient Descent (ASGD) to deal with the parameter trainingprocess instead of SGD so as to remove the inter-dependence between the item and userparameters.④Subsequently proposing a parallel new model named Parallel collaborativefiltering recommender based on Regularized Matrix Factorization(P-RMF), of whichthe training process can be parallelized through simultaneously training different user or item features, which is able to ensure the recommender accuracy, in the meanwhile,improve the operating efficiency.⑤Conducting experiments on two large and real datasets, say MovieLens1M andNetflix, and illustrating that the P-RMF model is capable of proving a faster solution tocollaborative filtering problem, as well as accurate recommender results whencompared to the original collaborative filtering based on regularized matrixfactorization and the parallel alternating least square recommender.