Network Measurement Based Quality Assessment Methods For Mobile Video Service

With the development of the mobile network and the increasing bandwidth,the traffic for video service becomes dominant and is still growing in the mobile network.Thus,it is critical for network operators to assess the quality of video service to ensure the quality of experience(QoE)and to manage and optimize the network.There are mainly two problems for network operators when assessing the quality from the network side:Firstly,the quality cannot be directly evaluated by the quality of service(QoS)parameters due to the application of HTTP adaptive streaming(HAS).Secondly,the adoption of end-to-end encryption techniques invalidates deep packet inspection(DPI).In this case,the network operator cannot assess video quality by parsing the content and even cannot identify the traffic for video streaming.Therefore,in the context of the application of HAS and end-to-end encryption,we focus on the critical issues of video service quality assessment from the perspective of mobile network operators,mainly to study the following aspects:Firstly,for identifying the traffic of HAS,we propose dynamic warping network(DWN)model based on soft-dynamic time warping.By extracting features of the traffic patterns,we first identify the traffic for video service in combination with traditional flow statistical features,then classify different video services into different categories,thereby completing the identification of the traffic for HAS.DWN takes the network layer QoS parameters as input to extract the features of the traffic patterns,which makes it possible to identify the traffic for HAS effectively and applicable to various encryption scenarios.Secondly,for the quality assessment of non-encrypted HAS,we propose a hybrid video service quality assessment method based on the player model and machine learning based methods.Machine learning(ML)algorithm is first used to assess the overall quality,i.e.,to assess if there exist quality issues in a coarse-grained manner,then the video identified with stalling events is further assessed using the player model based method by reconstructing the playback process.The quality is assessed in a fined grained manner,where the stalling occurrences and duration are both evaluated.The hybrid method reduces the number of the video that needs to be reconstructed by eliminating the video without quality issues,which reduces the complexity of the assessment while ensuring the performance,and also achieves to assess the quality in both coarse-grained and fine-grained manner.Thirdly,for assessing the overall quality of encrypted HAS,we propose a video quality assessment method based on the protocol independent features constructed from the network traffic.After estimating HAS segments from QoS parameters,the features of the player buffer are constructed.Combine with the statistical features of the segments,the overall quality of encrypted HAS is assessed using ML algorithms,where initial buffering is estimated,and three questions for stalling are answered:is there any stalling,are there multiple stallings and is the stalling ratio larger than 10%?The proposed method takes consideration of the characteristics of HAS,and also takes the advantages of the traditional player model based and statistical feature based methods to construct features from the time series of download speed,which ensure the effectiveness and the protocol independence.At last,for the fine-grained quality assessment of encrypted HAS,we propose attention based hidden Markov model(AHMM).The playback process is first modeled with HMM,where the time series of the download speed is taken as the observation sequence,and the playback state sequence is considered as the hidden state sequence.Then,we introduce and modify the attention mechanism to calculate the time-varying transition probabilities.AHMM is an end-to-end model,which takes the download speed series as input and outputs the playback state sequence by Viterbi decoding,and fine-grained quality assessment can be achieved by analyzing the output state sequence.By controlling the decoding depth of the Viterbi algorithm,AHMM can be applied to real-time or quasi-real-time scenarios.Based on the data collected from our self-developed data acquisition platform,this dissertation studies the critical issues of video service quality assessment in mobile network.By this study,we achieve to identify the traffic for encrypted HAS,to assess the quality for both non-encrypted and encrypted HAS in both coarse-grained and fine-grained manner,which make great sense to guarantee the QoE and to manage and optimize the network for mobile network operators.