Measurement And Prediction Of QoE Of A Large-scale Online Video Streaming System

With the rise of online video streaming services on the Internet,understanding and predicting user experience quality(QoE)has become particularly important.Accurately predicting the QoE problem of user sessions can help service providers take timely response strategies(eg,allocate special bandwidth resources),improve user QoE,so that increase user loyalty,which is of great significance.Video QoE prediction is the current research hotspot.At present,a large number of machine learning and deep learning models are used in the research community to solve this problem.However,accurate prediction is still challenging.This is because video QoE is affected by many factors,but it also changes over time.In addition,another key challenge that affects the accuracy of prediction is the sparseness of the feature space,that is,the value of the attribute characteristics of the actual video conversation is too much,resulting in too large feature space.There are only sparse training samples in the feature space,so it is difficult to find similar sessions to predict the QoE of the target sessions.The work and contributions of this article for the QoE prediction problem of video are as follows:(1)Based on a large-scale,user-viewed QoE measurement data from a real-world online video system,the problem of feature space sparseness for session-level QoE prediction is analyzed.The complex characteristics of user QoE changes are measured and revealed,and various relevance of session characteristics,including ISP(Internet Service Provider),user location(province),video popularity,video type and user viewing habits,we found that statistical characteristics(mean,median)of QoE aggregated based on these characteristics are not only stable,but have different patterns of change over a relatively long period of time,so it can be used as a feature to predict QoE.(2)A novel and effective feature characterization method is presented.The above features are characterized and used as the input to the QoE prediction model.This solves the problem of sparseness in the feature space for session-level QoE prediction.The method is that instead of using attribute values(eg,video ID,ISP ID,location ID,and user ID)of each feature as the model input,but using the QoE statistics feature that satisfies the set of sessions under each feature condition as input.For example,in order to predict a Beijing user to watch QoE of video A through Unicom,we do not use the user's location ID(Beijing),ISP ID(unicom),and video ID(A)as model inputs,but instead separately calculate the average and quartile of QoE of sessions in Beijing,Unicom,and video A as input to the models.Experimental results show that this feature representation method achieves 23.26%higher performance gain than the traditional feature representation method.(3)Comprehensive use classification trees and probability inference algorithms to analyze the relevant factors affecting users' QoE viewing video,understand the degree and characteristics of each feature's impact on it,and find that the most influential feature is the video name + location(province)QoE,which can help provides actual QoE optimization in the system a reference direction.(4)Build and evaluate various machine learning and deep learning models to predict session-level video QoE.The experimental results show that our machine learning model can obtain stable prediction accuracy of 67%with the help of our feature representation method.