The Research On Personalized Recommendation System Based On Bipartite Network Configuration

With the rapid development of information technology and the coming of the era of "Internet plus" and big data, the phenomena of information overload have deteriorated. It is an emerging issue concerned by scholars that when the information is needed and how these resources around us can be gathered, synthesized, analyzed, evaluated and predicted. The recommendation system based on bipartite networks builds a good platform to solve the above problems, which has been applied in many engineering fields.In addition to the convenience brought by the recommendation system, there are also some problems:on one hand, there is a lack of evaluation enthusiasm of users, as a result of sparsely relative data. On the other hand, malicious scoring to some projects by some users results in the declination of the credibility of the recommendation system, threatening the system security and justice. Furthermore, it will affect the accuracy of recommendation.In this paper, we try to improve the problems in the bipartite network nowadays depending on the analysis of the domestic and foreign research. According to the users’rating database, community of users of average scores is constructed and traditional credibility algorithm is used to calculate the initial credibility of the users. With the "average" users’ reliability value as a benchmark, deep learning algorithm for the second assessment of the credibility is applied to other users, the results of which are arranged in the ascending order. Suspicious users ranking top-L will be removed to create a trustfully adjacent collection for the target user, which enhances the accuracy of the recommendation system.With sparse statistics, traditional bipartite networks recommendation algorithms have deviations when calculate the similarity between users and recommendate for the target users. At the same time, the weighted score division is not accurate and meticulous enough. This paper puts forward to a new recommendation algorithm based on the monotonous saturation function, and takes the tangent of target users and the other projects’common rates number against all users as the traditional similarity coefficient. At the same time, after the coefficients get adjusted, the similarity will be placed in descending order, and the set of the former K-nearest credited collection can be recommended for target users. The experimental results prove that the revised algorithm has both improved the recommendation in accuracy and reduced the complexity.Finally, the paper designs the recommendation system of books, including the login module, the credibility evaluation module, the score message module and the recommendation module. In the premise of the analysis of users’ background and interests, the system can recommend bibliography to target users by using the adjacent collection or keywords that the users give. Before recommendation, the system will block the suspect users according to the credibility evaluation in order to safeguard the fairness of the recommendation system and to improve the user experience.