Research On Recommendation Methods Based On Trust Perception Of User Behavior

Recommendations provide great convenience for users to use Internet.According to users' input information,recommendations can be divided into(1)keyword based recommendation,(2)potential user behavior and relationship based recommendation.The most commonly used keyword based recommendation is search.The process of search is that a user enters a keyword and the search engine returns links most similar with the keyword could be regard as a recommendation with clear goals.However,the traditional search engines try to satisfy users by improving the recall and precision of search results,which ignore the distinction between different users in searching the same keyword.Improvement of recall and precision of search results can not meet the personalized needs of users.Search engines depend on keyword can not satisfy personalized search that is caused by the limited information of users or lacking users' implicit demands.One of the key factors to improve search quality is to meet the personalized needs of users.Personalized search is mainly used for users with different backgrounds or different expectations.The quality of personalized search depends mainly on the accuracy of mining users' implicit information and reflection of users' real-time changes of preferences.Based on systematic research on search behavior,this paper puts forward improving search quality from the following aspects:(1)An implicit prediction method for personalized search is proposed,which is based on user behaviors.In this method,history of user's behaviors is used to reflect user's preference.A hidden Markov model is built to describe the relationship between user's behaviors and preferences,and is used to predicted user's search preferences.The proposed method could give personalized search results with less time complexity.(2)Iillegal rasing PageRank of a webpage has an adverse effect on improvement of search result quality.For a webpage with high PageRank will appear in the front of the searching results,the ranking of webpages affect the search results.In order to rank in the top of the searching results,many webpages are illegal improved by artificial webpages.In this paper,abnormal webpage ranking detection method is proposed to improve the search result quality.Another issue of recommendation is how to provide reasonable information for users who can not give explicit keyword.In this paper,user behavior and perception of trust are studied to improve recommendations.Through the study of user behavior and trust relationship,this paper puts forward the recommendation quality from the following aspects:(1)Research how to provide reasonable recommendations for users who are newly join in,namely user's cold start problem.For users trusted by others give more credible recommendations,in this paper,a method is proposed to constrain propagation of trust relationship by distrust,which also combines the information of user-item matrix with expanded trust relationship to improve the effectiveness of recommendations for new comers and users with little records.(2)Investigate the effectiveness of time on the recommendation.After researching the relationships between user behaviors and their preferences,a time-sensitive model is proposed to follow the changes of user's preference.In order to reduce the computational complexity of user similarity,user similarity calculating is cast into optimal matching of bipartite graph,which not only ensures the CF algorithm's accuracy,but also reduces the time complexity.(3)In order to resolve the issue that people pay little attention to unpopular commodities,a recommendation method is proposed to give recommendations constrained by long tail distribution.In this paper,user similarity is found by their behaviors and is expanded with the constraint of distrust.The proposed method promotes the weight of unpopular commodities,and resolves the issue that users pay little attention to unpopular commodities or goods with few comments.