Modeling And Optimizing Quality Of User Experience For Personalized Online Video System

Online video service has become the mainstream application of the Internet.In order to satisfy users and benefit video service providers,how to understand,predict and improve the quality of user experience(QoE)in online video services has become a common concern of researchers.On the one hand,in order to accurately predict QoE,the existing work mainly focuses on the relationship between QoE and objective quality of service(QoS)factors such as throughput and buffer rate.However,in fact,the subjective user preference on video content is also closely related to QoE.Moreover,a certain degree of user preference and QoS level may also trigger user interaction during the viewing process,such as seek and pause,which will also affect QoE.To better understand and predict QoE,the following challenges still need to be addressed:how to obtain user preferences that cannot be directly observed;how to analyze and model the mapping relationship between QoE and the above factors.On the other hand,the existing QoE optimization methods mainly adapt to users'needs through network resource scheduling.However,with the richness of video content and the increase of personalized preferences,the distribution of video demand is becoming more uneven,and the long tail effect is becoming more serious.These methods have been difficult to cope with the challenge that the network resources serving for the cold video are wasteful and those for the popular video are insufficient.In view of the above challenges,this paper carries out the following research:1.Based on large-scale real data,mining personalized user preferences from a few viewing records;2.Analyzing the relationship between QoE and user video preferences,QoS and interaction behavior,and accordingly establishing a QoE predictive model;3.From a new optimization perspective,proposing to maximize the average QoE of all the users in the system by guiding the user's needs to achieve the global optimal configuration of network resources and user needs.The specific work and contributions are as follows:Firstly,this paper proves that users' viewing behaviors are influenced by many factors and can not be directly regarded to be user preferences.A user preference extraction-inference framework is proposed.Based on a small amount of behavior data which can truly reflect user's preferences,the user's personalized preferences for all videos are inferred.Experiments show that compared with the direct use of all data,the proposed frame improves the inference accuracy by 4.8%.Furthermore,through measurement,this paper proves that users have their personalized preferences on video popularity and accordingly proposes an improved preference inference algorithm.Furthermore,this paper also proposes correction algorithm based on collaborative filtering to reduce the statistical bias on user popularity preference caused by data sparsity.Experiments show that the proposed algorithms can not only improve the inference accuracy in sparse scenarios,but also improve the TopN recommendation coverage without any other data sources.Secondly,through measuring and analyzing the relationship between QoE and QoS,user preferences and interaction behavior,two sets of QoE predictive models are established.It is found that the user's sensitivity to the degrade or upgrade of QoS varies with her preference degree,and the interaction behavior such as seeking may not only affects QoE,but also has some correlation with user preference and QoS.Therefore,this paper constructs a nonlinear QoE prediction model based on QoS and user preference,and a seeking behavior predictive model as well as a hierarchical QoE predictive model with seeking behavior as an intermediate variable.Experiments show that compared with the traditional QoS-based QoE model,the proposed models can greatly improve the prediction accuracy,which confirms the importance of user preferences and interaction behavior for accurately understanding QoE.The proposed models are in analytical expressions,which is helpful for quantitative QoE optimization.Thirdly,this paper proposes a new recommendation-based QoE optimization algorithm based on queuing theory and QoE predictive model.To improve the global QoE in the system,the proposed optimal algorithm guides users' video selection behaviors in different scenarios.Different from the traditional optimization methods such as resource scheduling,this algorithm takes into account the impact of both video content and QoS on QoE,as well as the impact of recommendation results on user behaviors.Simulation results show that the proposed algorithm can improve the global QoE under certain bandwidth resources,compared with traditional recommendation strategy,uniform allocation strategy and heuristic QoE greedy strategy.The more unbalanced the video popularities,the more significant the advance of the algorithm.