Research On Personalized Recommendation Methods Based On Social Tagging System

Recommender systems are an important part of modern information systems,and have been hot research topics in machine learning and information retrieval.With the development of Web2.0,social tagging systems have been widely used because of their sharing,descriptive,and interactive features.Compared with the traditional recommender system,the social tagging system can use both social tagging and rating information for recommendation,which greatly improves the interpretability and accuracy of the recommendation results.However,the recommendation technology based on the social tagging system is still in the initial stage of development.The traditional recommendation method has certain limitations in the new context of the social tagging system,which is mainly reflected in:(1)the inability to effectively perceive changes in user interests based on social labeling and rating information within the scope of social tagging systems;(2)The social characteristics of the social tagging system and the multi-granularity of user interest topics cannot be effectively used to accurately discover the user interest communities;(3)The social tagging and rating information cannot be effectively integrated and personalized recommendations can be made.This paper aims to improve the quality of personalized recommendations in social tagging systems,and uses related theories and methods such as topic models,state space models,and matrix factorization to study the problem of personalized recommendation in the context of social tagging systems.The personalized recommendation theory and method of the social tagging system enhances the practicability and flexibility of the personalized recommendation algorithm.Specifically,this article has carried out the following research:(1)Research on user interest drift perception methods.Aiming at the problem of calculating the entire training data as a whole without distinction of time in the current social tagging system and ignoring user interest drift,a user interest topic drift perception method based on a dynamic topic model and a state space model and probability matrix decomposition are proposed.Fusion user interest level drift perception method.The proposed method effectively utilizes social tags and rating information to identify the user's interest topics and the life cycle of the user's degree of interest,so that the periodic characteristics of user interest can be effectively obtained.(2)Research on multi-granularity interest community discovery methods.Aiming at the problem that the user interest community detection in the current social tagging system does not consider the difference between the interest topic and the tag attribute,which leads to the inability to effectively discover the user interest community,a multi-granularity interest community detection based on non-negative matrix factorization and nonlinear matrix factorization.The proposed method uses non-negative matrix factorization and nonlinear matrix factorization methods to identify the user's granularity of interest,so that it can discover and identify communities of interest from the multi-granularity levels of user interest topics and tag attributes.(3)Research on personalized recommendation method based on fusion of social tags and rating information.Aiming at the problem that the current recommendation methods in social tagging systems do not effectively integrate social tagging and rating information,a recommendation algorithm based on the fusion of latent Dirichlet allocation and probability matrix decomposition is proposed from the perspective of user interest topic level and tag attribute level,and joint probability is used.Recommended algorithm for matrix factorization.The proposed algorithm effectively fuses social tags that reflect the user's topics of interest with rating information that reflects the user's degree of preference,which can effectively solve the problem of unclear semantics of rating information and the matching of rating and item attributes.(4)On the basis of the theory and method proposed in this paper,a personalized recommendation system for online product sales is designed and implemented.The system implements various algorithms proposed in this article.The data analysis function of this part is structurally independent.For business modules,it has better portability and maintainability.