Research Of Hybrid Collaborative Filtering Recommender Algorithm Based On The Information Of Features And Preference

The Internet has become an indispensable tool for human development. Peoplecan obtain critical information needed in order to solve the basic necessities of dailythrough the Internet. But the impact of Internet information overload has become anincreasingly serious problem. The main methods of solving the problem ofinformation overload are portals, categories, search engine and recommendationsystem.Collaborative filtering algorithm gets the widest range of recommendedalgorithms. Nonetheless, data sparse, cold start, scalability, and security issues haveprevented further application for it. At the same time, group features information(such as user profile and project feature) can solve sparse and cold start problems.First, this paper introduces the research status of the recommendation system.Second, the paper also analyzes and summarizes advantages and disadvantages aboutcollaborative filtering algorithm. At the same the article get different categories ofrecommendation algorithms. On the basis of research results, we have some effectivesolutions for traditional collaborative filtering problems on the accuracy and datasparse.The principal contents consist of these two aspects.1. For information on group identity, this article features information for usersand items was extracted using feature information to build a user-item characteristicsrating matrix and a user characteristics-item rating matrix. Furthermore,we optimizethe characteristic matrixes. Users’ features, such as age, sex and residential areas, etc.,and features of the items’, such as their type, shelves and time, are constructed to theuser characteristic-item rating matrix and user-item features scoring matrix.Meanwhile, in order to enhance the recommendation accuracy, we use hidden featuressemantic model LFM to optimize the feature matrixes. The reason for using the LFM model optimization rather than using SVD is that LFM model can get the best suitablefor the actual situation.2. For the problems, such as the similarity of user similarity and item similarityare not high, computing time and space overhead is too large and so on, we use theuser characteristic-item rating matrix and the user-item characteristics matrix toscoring matrix similarity similarity between users and items. More, we use twosimilarity fusion methods (a simple linear combination method, and the other way isthe integration matrix multiplication) to fusing the similarity from the user-itemrating matrix and get a new similarity matrix. Finally, the classic formula forcollaborative filtering scores prediction score prediction.The experiments prove that the proposed recommendation algorithm based onhybrid feature’s reference information to improve prediction accuracy score has apositive effect.