A Research And Optimization Of Learning To Rank Based Personalized Recommendation Algorithms

Recommender system(RS)is an effective tool to tackle the problem of information overload on the Internet.Conventional recommender algorithms always focus on the accuracy of items' prediction scores without considering items' rank in recommendation list.However,a good accuracy of prediction scores does not ensure a good ranking performance for users.In order to get a better ranking result,some researchers try to integrate learning to rank(LTR)methods into recommender systems to construct ranking-based recommendation models.In this thesis we have studied how to use machine learning method to improve the ranking performance of recommendation algorithm.Within this thesis we firstly introduced the conception of RS as well as some common recommendation algorithms.We also introduced some LTR methods which were used in the field of RS.Then we proposed two novel ranking-based recommendation algorithms to handle with implicit feedback data on the basis of analyzing several current learning to rank methods.In the first proposed approach,we focused on the problem that most traditional recommendation algorithms only concentrate on predicting single item's score without considering the order relationship between recommended items.We proposed a pairwise ranking method,named pairAUC,which utilize the items' position information adequately to improve the ranking performance and used the relative ranking positions of a pair of items to control the learning step size of the model parameters.By optimizing the loss function related to AUC to get the ranking-based recommendation model.Then we focused on the problem that pairwise method didn't consider the item's rank position in recommendation list.Combining with the listwise ideology we presented the second approach that train the recommendation model by directly optimizing the ranking metric MAP,named MSMF.We optimized the metric by using smoothed approximation functions instead of the non-continuous parts of original MAP.For further improving the rank performance of MSMF,we proposed AdaMSMF on the basis of MSMF combined with the thought of AdaRank.This method created a strong ranker by composing some weak rankers linearly with weight function to make recommendation.Experimental results on different datasets showed that our proposed pairAUC model outperformed traditional pairwise recommendation algorithms like BPR,FISM and RankBPR.And both our proposed MSMF and AdaMSMF outperformed the state-of-art listwise LTR algorithms like CLiMF,xCLiMF and LRMF.