The Study On Recommendation Systems Of Smart TV Based On A Time-aware Bayesian Ranking Model

In recent years,wit.h the rapid popularization of network technology,smart,TVs have gradually replaced traditional cable TVs into thousands of household-s.Based on web technology,smart TVs enable viewers t.o easily access massive amounts of the internet programs on a single platform.Different from the pro-grams in the cable TVs where users can only watch t.he program broadcasting on current TV station,on smart TV platform users can select any program with permission on the network to watch.At the same time,a large amount of TV pro-grams on the Int,ernet,make it difficult,for users to select their favorite programs.In addition,smart TV program recommendations are more challenging than con-ventional video recommendations.On the one hand,smart TVs have the problem of shared account,which is multiple family members share a common smart TV account.On the other hand,the problem of weak implicit feedback exists in the viewing behaviors of smart TVs,which is not all viewed programs should be regarded as positive feedback.In tradit,ional implicit feedback recommenda-tion systems,user-interactive items are generally regarded as positive feedback.However,in smart TV scenario many programs are viewed just for a while by users during exploring prefe.rences,and these programs may not be.preferred by users.Therefore,it is crucial to address above two challenges and recommend personalized smart TV programs to users based on their viewing history.As the largest,TV manufacturer in China and the third largest TV manu-facturer in the world,Hisense has tens of millions of users.To address these new challenges,we analyze user viewing records on Hisense s smart,TVs,and find three interesting phenomena.The first phenomenon is that different fam-ily members have different.program viewing time.The second phenomonenon is that programs viewed by a family in adjacent time slots are similar.The third phenomenon is that program viewing durations reflect user preferences.We then conduct statistical analysis to verify the three phenomena,respectively.Based on these phenomena,we propose three mechanisms to improve recommendation per-formance.Specifically,we divide one day int.o six time slots by K-means based on Phenomenon 1.Thus different,family members' viewing behaviors roughly correspond to different time slots,which can be used to relieve the problem of shared account in smart TVs.To address the problem raised by Phenomenon 2,we introduce the sliding window method and regard each time slot and its adja-cent time slots as a window.We present,an intuitive strategy to map the weak implicit feedback to explicit scores based on Phenomenon 3,which can be used to address the problem of weak implicit feedback in s,mart TVs.Finally,based on Bayesian Personalized Ranking(BPR)framework,we formalize Time-aware Bayesian Personalized Ranking Approach(TABPR)as a tensor model by intro-ducing time slots as additional dimensions.And we try to learn different user preferences in different time slots and meanwhile solve the problems of shared account and weak implicit feedbackExtensive experiments on a large-scale real-world dataset,the Hisense smart TV dataset,demonstrate the effectiveness of our TABPR.And it has significant improvements than the state-of-the-art methods.Further analysis experiments show that all mechanisms based on above three phenomena can improve recom-mendation performance,demonstrating the effectiveness for all.