Research On The Application Of Learning To Rank In Personalized Recommendation Systems

The research of personalized recommendation system has been great concernedwith the development of e-commerce in recent years, and it plays an increasingimportant role with the fields of internet application further expand. However, mosttraditional recommendation algorithms are registered user-facing or the scope of itsapplicability is limited, on the other hand, with the dramatic rise in the number ofinternet users and greater variety of internet services, new recommendation algorithmsare strong demanded to meet different requirements. What’s more, each application fieldalso needs special algorithm because of its own characteristic. Starting from these actualrequirements, this thesis carries research on unregistered user-facing recommendationalgorithms first.In traditional recommendation system, unregistered users tend to can’t acquirerecommendation service or the quality is poor, this thesis proposed two algorithmsspecifically for such user, which extended from content-based recommendation andassociation rule-based recommendation. These algorithms support by informationextracted from many scattered records unregistered users left after browsed. The testsindicated that these algorithms had yielded good results in certain areas, but insufficientavailable information limited their application.Most recommendation systems predict user’s preference by current browsingcontent and access log, but for many application fields in the information age, thepage’s freshness, click rate and other users’ browsing performance maybe moreimportant than personal preference since user’s interest will changes as eventsdeveloping and time passing in those fields, traditional recommendation system can’treflects the change timely because of ignore these factors.The thesis propose an innovative idea of feedback on the effectiveness for theseproblems, every user access will not only be recorded as access log, but also feedback tothe recommendation system. And for the target object to be recommended, the predictedrating is no longer changeless, except for affected by traditional factors such as querysimilarity, content-length, PageRank and number of links, it also be determined by content-freshness, click rate and other users’ evaluations, all of these factor will nodoubt increasing the objectivity of the comprehensive evaluation by all users, and theevaluation in turn will contributed to predict personal preference.How applying the idea of feedback to specific recommendation model is anotherimportant content of this research, learning to rank is a way to solve the problem, thisthesis discusses how applies it to traditional recommendation algorithms in detail, andtests its effectiveness and adaptability by experiments. The results show that, comparedto traditional recommendation model, the feedback recommendation model can improvethe accuracy of prediction ratings effectively on the premise of not significantly degradethe system performance.