Research On Collaborative Filtering Recommendation Algorithm Based On Joint Learning Of Tags And Ratings

popularity of the Internet have greatly promoted the development of the E-commerce. It brings great conveniences to people’s production and life, but causes the problem of data overload at the same time. Recommendation system is one of the effective way to resolve this problem. It can simulate the behavior of shopkeepers to help customers finish the process of buying and save time for users while searching for products. Meanwhile, a recommendation system can also help the E-commerce providers to quickly find their potential customers and then improve their profits. Therefore, recommendation system has become the hot research issue of domestic and foreign scholars. The recommendation system utilizes the historical behaviors of users to infertheir interest models. And then it selects appropriate objects from the massive data set to make recommendations for users. Recommendation algorithm is the key part of a recommendation system. Recommendation algorithm mainly includes content-based recommendation algorithm, collaborative filtering recommendation algorithm, structure-based recommendation algorithm and hybrid recommendation algorithm that integrates different recommendation algorithms. Among these algorithms, collaborative filtering algorithm is one of the mainstreaming recommendation algorithms. It has been widely used in some commercial recommendation systems. Many researchers focused on the rating-based collaborative filtering algorithm and performed in-depth study to it. And they have made significant progress. But due to the problems of data sparsity, cold-start and so on, the accuracy of recommendation algorithms remains to be further improved. Since similarity measurement of users and items is the main factor of affecting the quality of recommendations, looking for auxiliary information, such as tags information, and taking advantage of them to improve similarity measurement becomes one of the tasks of researchers. Therefore, for rating data, we research around the following two aspects:(1)Designing of user(or item) similarity measurement strategy; (2)How to employ tags information effectively.1. Propose a new collaborative filtering recommendation algorithm thatintegrating both tags and ratings(UTR-CF), in which we employ tags information and ratings data simultaneously to design new user(or item) similarity measurement strategy. It can overcome the shortcoming of utilizing ratings data only for similarity measurement. The experimental results on MovieLens data set show that the UTR-CF algorithm can improve the recommendation accuracy effectively and solve the issue of data sparsity to some extent.2. Propose a new collaborative filtering recommendation algorithm that embeds the LDA topic model(ULR-CF), in which we utilize LDA topic model to mine latent topic information to obtain document-topic distribution at first. Furthermore, we design a new similarity measurement strategy based on document-topic distribution and ratings data. The experimental results on MovieLens data set show that the ULR-CF algorithm can overcome the problem of data sparsity and improve the recommendation accuracy effectively.3. Propose a new rating matrix factorization algorithm that integrates local similarity(ELS-MF). In order to further improve the performance of the existing matrix factorization algorithm, a local similarity constraint term is embedded into the square error minimization function. Along this way, we can get more accurate latent vectors in latent space, so more accurate recommendation results can be obtained. The experimental results on MovieLens data set show that the ELS-MF algorithm can reduce the recommendation errors of recommendation system effectively.