Font Size: a A A

Study On Quality Of Experience (QoE) Of Internet Video

Posted on:2022-01-22Degree:DoctorType:Dissertation
Country:ChinaCandidate:T YueFull Text:PDF
GTID:1488306326979389Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Internet video is currently one of the most popular Internet services.Improving quality of experience(QoE)is increasingly significant for service providers,since bad user experience gives rise to customer churn rate.QoE-oriented service management can effectively improve the user experience while saving operating costs.Studying Q oE in the real world is challenging.Firstly,influence factors(IFs)of QoE are various and distributed in different dimensions.Features corresponding to IFs are heterogeneous.Secondly,there are multi-order interactions that exist among multi-dimensional features.The correlations between features and QoE are nonlinear and complex.Last but not least,session-based user behaviors are related to QoE and of benefit to QoE management,but the temporal and interactive property,as well as individuality of user behaviors make the issue more difficult.Based on the massive user data collected from the client-side players of a large-scale Internet video provider,this thesis presents a data-driven and comprehensive study on QoE of Internet video.The main works include QoE measurement and learning,QoE evaluation modeling and QoE management and optimization.The contributions of this thesis are listed as follows.1.QoE measurement and learning is the basis of this paper.By building a system of data acquisition and preprocessing on the client side,we extract features of different types according to QoE IFs in three dimensions:system,context and user.Video sessions are described by multi-dimensional features.For system aspect,we measure user-perceived quality of service(QoS)metrics in application layer that are affected by media-related and transmission-related IFs.For context aspect,we consider the IFs both from server and terminal,and propose to introduce context IFs as a sub-vector in feature space for QoE modeling,considering the extension of the sub-vector.For user aspect,we introduce pair-wise pattern features of session-based user behaviors,considering the temporal property of user behaviors and the interactions between user and system.A novel algorithm of frequent time series pattern mining is proposed,which is called lexicographic hierarchical intersection(LHI),aiming to capture the pair-wise patterns for user feature extraction.2.QoE evaluation model is the vital issue of this paper.Aiming to learn the multi-order interactions among features on the video session set with the sparse and high-dimensional space,we present a QoE evaluation model with feature engineering and a non-linear ensemble algorithm,which maps multi-dimensional features to one QoE function.With the method of feature engineering,we apply fuzzy theory to normalize continuous features of uneven value distribution.Verifying the positive correlation between the support rate and discriminative capability of typical pattern features.And an ensemble approach is developed which is called cascaded bagging-based Bayesian factorization machine(CB-BFM).CB-BFM utilizes latent factors to model interactions among features in the absence of interaction information.Reducing the negative impact of meaningless interactions both on effectiveness and efficiency of modeling.The experiments verify the effectiveness on choosing basis classifier,introducing user behavior pattern features and applying ensemble approach.3.In order to solve the problem that the data-specific methods of feature engineering have an adverse effect on model generalization,we conclude QoE evaluation modeling as an issue of representing learning on features of different types further,developing a deep learning based model for QoE evaluation.We present a hybrid deep network that integrates a deep neural network(DNN)and an improved recurrent neural network(RNN),called DIRNN.By optimizing the network,DIRNN conducts representation learning both on sequential and non-sequential features of three dimensions simultaneously.Contextual information and time difference are incorporated to different layers of RNN.Attention mechanism is applied for further improvement of DIRNN.The experiments demonstrate that DIRNN achieves better than related works of QoE evaluation and some typical deep networks of another applications of Internet.4.Based on the study of user behavior on the video session set,we explore a fine-grained approach of QoE management and optimization,with the help of QoE evaluation model.The objective is to prevent QoE from degrading continuously by monitoring and intervening with user behavior during video session,and then improve the user engagement.Firstly,we conduct a graph-based ranking algorithm on the basic of attention mechanism in a supervised way.After filtering noise of time sequences of video sessions,we obtain typical user behavior patterns that are related to QoE positively and negatively.And we obtain expert knowledge about typical behavior patterns via studying on the video sessions.Then we propose a control mechanism of user behavior monitoring and intervention(UBMI)at client side,the internal logic is based on the expert knowledge.By interacting with QoE evaluation module,UBMI monitors user behaviors and update decisions during video session continuously.UBMI identifies potential negative behavior patterns,and decides to intervene according to real-time state of player and network.
Keywords/Search Tags:QoE, feature engineering, ensemble algorithm, deep learning, user behavior
PDF Full Text Request
Related items