Research And Implmentation Of Hybrid Recommender System Based On Bipartite Network

Recent years, with the fast development of Internet and E-commerce, the scale of the information on the Internet grows rapidly; too much useless information makes users difficult to find items or information that they need. In this situation, personalized recommender systems have been proposed to solve this problem.Personalized recommender system can recommend items and information to users according to users’interests so that users can find what they really interest in with little cost of time. But with the widely use of personalized recommender system, some new problems are raised, for example cold start and data sparse problems, recommender system based on a simple recommender algorithm cannot handle these problems. At the same time, recommender system based on serial algorithms cannot deal with the larger scale of data on the Internet.According to problems mentioned above, in this paper, we have researched hybrid recommender system based on bipartite network. At first, we have introduced some basic concepts of recommender system. And then, latent factor model methods and recommender algorithms based on bipartite network are introduced, and a latent factor model method called SLIM (sparse linear method) is described particularly. We analysis the implementation and key idea of SLIM and find that SLIM has a limitation:similarity of items that have not been co-purchased by any user cannot be studied by SLIM, then we provide a new hybrid recommender algorithm called UIIM (user-item interest method) based on bipartite network to improve the performance of SLIM. In order to handle the rapid growth of data, we implement the parallel UIIM method called Parallel_UIIM based on Spark. The experiment shows that UIIM achieves a better performance and recommender quality than SLIM, and for users that have few purchasing history, UIIM also outperforms over SLIM, that indicates the problem brought by data sparse can be improved by UIIM. And experiment also shows that parallel implementation of UIIM based on Spark can deal with the larger scale of data, and the parallel algorithm Parallel_UIIM outperforms over serial UIIM algorithm with large-scale data sets. At last, we design the architecture based on massive data analysis platform, then we also implement three algorithms based on Mapreduce for data pre-processing. We build a hybrid recommender system based on hybrid algorithm UIIM on massive data analysis platform; this provides a sample of hybrid recommender system for practical application.