On Predicting Quality Of Experience In Online Video Streaming Services

With the continuous development of Internet technology and video multimedia technology,online video has become an important way of leisure and entertainment.Cisco's Internet forecast report showed that:in 2015,video traffic accounted for 70%of total Internet traffic,and video traffic will account for 82%of all the consumption network traffic by 2020,of which mobile video data traffic will account for 50%.Such a huge video data traffic has brought great challenges on the current online video streaming services,especially mobile video services.Meanwhile,video users also made a higher level requirement on the quality of video viewing:high video resolution,low start-up delay,low buffer rate.They pursue a higher quality of experience(QoE).Therefore,how to predict and improve quality of experience accurately in network video services,has great theoretical value and commercial value.The existing studys on quality of experience,mostly study video user viewing behaviors and video quality impact factors,or put forward some complex control platform systems to optimize the transmission efficiency of network video resources,or study complex video encoding methods,to improve quality of experience finally.We intend to use machine learning algorithm to construct a simple and easy-to-deploy QoE model,which based on user terminals,to improve quality of experience.This paper has four specific contributions as follows.(1)We have analyzed the PPTV video user access log dataset in detail and found that 1)the initial buffer length needs more targeted optimization than total buffer length;2)the number of buffer events has the greatest correlation with user's valid viewing time ratio.Based on these results,we designed a high performance QoE mapping model based on random forest algorithm.Our model was accurate that F1 score of user QoE-bad prediction reached 0.77.Moreover,the initial buffer length and the number of buffer events have more influence on prediction performance of QoE mapping model than other features.(2)We have developed a complete set of experimental platform for DASH-based video quality research in LTE network environment.Specifically,we deployed DASH video server and MongoDB database on the Ali cloud servers,developed an Android application for measuring LTE network quality parameters,and modified the dash.js source code for capturing playback information of DASH video client.(3)Through analyzing the experimental measurement data statistically,we found that video freezing event will occur when the buffer length is below 0.5 second.And the current DASH implementation's QoE problems were primarily due to round trip time.(4)We proposed a prediction method based on "time window",and designed two QoE models based on random forest algorithm.Our models were accurate that F1 score of user QoE-bad prediction reached 0.87.Otherwise,proper size of interval time window is 28 seconds,and proper size of history time window is between 10 seconds and 18 seconds.