Research On Parallel Algorithm Of Collaborative Filtering Based On Matrix Factorization

With the rapid development of Web2.0 and Internet-based business, in the field of e-commerce, there is an increasing number in amount and types of commodities. Moreover, the online to offline mode has been widely applied in s uch area. However, it is must be admitted the fact that customers have to spend so much more time to seek their preferences that countless of time has been wasted. The process of browsing the irrelevant information and product will lose customers. So, the concern how to provide more accurate recommendations to potential customers has became more remarkable problem for platform of e-commerce.Nowdays, researchers prefer to put more attentions on enhancing the accuracy of recommdation without considering efficiency of executing process. Parallelization provides an effective way to improve efficiency of algorithms so that the research on parallel approaches of recommendation should be more focused. The dissertation intends to boost performance of collaborative filtering recommendation algorithm based on matrix factorization through parallelization and accelerated strategy with GPU. The main works of this investigation are described as follow:First and foremost, to make a solid foundation for further works, thi s dissertation investigates and analyzes mainstream collaborative filtering algorithms. The dissertation focus on researching the features and optimizing on technique of CUDA programming based on GPU. Then, the last step of work provides reliable support to implement of matrix decomposition on CUDA.In addition, the dissertation focuses on analysis of the recommendation algorithm of collaborative filtering based on matrix decomposition with the characteristic of high accuracy, the evolutionary model, the ap proach of training, analysis of theory and deduction of expressions et al. And the strategy and impact cause of parallelization are also discussed at the same time. As the result of this section, a new parallel algorithm, GPUMF, is proposed.Finally, the GPUMF Algorithm is implemented by the efficient GPU hardware and the frame of CUDA programming. The first group experiment in the thesis, with Movielens 10 M and Baidu Movie dataset and so on, proves that GPUMF has the ability to improve efficiency of executing significantly without damnifying accuracy. The second group experiment, whether the distribution of data affect the performance of GPUMF, indicates a balanced distribution can lead the relatively greater enhancement.