Personalized Recommednder System Based On Implicit Feedback: Resaerch And Implementation

With the fast development of Internet and Information technology, the usefulinformation resources that people can get increase rapidly. But in the meanwhile,people get into the trouble of “information overload”. Personalized recommendersystem can be able to push suitable online news to reader according to his interest soit can help people overcome the problem of “information overload”. In this thesis, wepropose a personalized recommender system based on implicit feedback and do someresearch on the technology of information filtering and the mechanism of userfeedback, which are key technologies of recommender system. The thesis focuses onthe following aspects:Firstly, the status of research on personalized recommender system is reviewedand the technologies of information filtering and the mechanism of user feedbackwhich are common used in current recommender systems are also introduced. In theaspect of information filtering, most of the personalized recommender systems nowuse the improved classification algorithm. And in the aspect of the mechanism ofgetting user feedback, most personalized recommender system now require user togive feedback explicitly. Explicit feedback forces user to engage in additionalactivities beyond his normal reading behaviors, and the benefits are not alwaysapparent. Combined with the results of former research, a framework of personalizedrecommender system base on user implicit feedback is proposed.Secondly, the information filtering technology which is one of the maintechnologies of personalized recommendation is further studied in this paper. Likemost of the personalized recommender systems, the improved Naive Bayesclassification algorithm is used in the module of information filtering. Currently,Naive Bayes classification algorithm can be improved in three ways: structureextension, attribute weighting and feature selection. But all these three ways willcause new problems such as computational-complexity increases. Based on deducingboth the expressions of attribute weighting and feature selection, a method throughadjusting the output of classifier is proposed. The improved classifier is used in themodule of information filter. The core idea of this method is adding a factor to theoutput of Naive Bayes classifier, and influencing the category of news by adjustingthe value of the factor. The value of the factor is adjusted by user feedback. Tests onseveral text datasets demonstrate that the performance of improved Naive Bayes classifier is better than that of classic Naive Bayes classifier.Finally, the mechanism of user feedback which is another key technology ofpersonalized recommendation is discussed in this paper. Because explicit feedbackforces user to engage in additional activities beyond his normal reading behaviors sothe benefits are not always apparent. Some works have showed that there is a closerelationship between user reading behaviors and interest. Based on these formerworks, a method which to get user interest implicitly by analyzing user behavior isproposed. First, user behavior model is described in Hidden Markov Model. Then thesequence of his reading behavior is analyzed with his behavior model, and his intereststate is gotten which is treated as feedback. Finally, user behavior sequence is used toupdate his behavior model. In order to collect data, a RSS reader is developed andsome students are invited to participate in the experiment. In the experiment, users arerequired to read online news with the RSS reader. The reader will record user readingbehavior and his explicit feedback. Tests on the datasets we collected demonstrate thatuser interest can be predicted with an acceptable accuracythrough behavior analysis.