No-reference QoE Prediction Model for Video
Streaming Service in 3G Networks
Xin Yu
1
Huifang Chen
1,2
Wendao Zhao
1,2
Lei Xie
1,2
1
Dept. of Information Science and Electronic Engineering, Zhejiang University
2
Zhejiang Provincial Key Laboratory of Information Network Technology
No. 38, Zheda Road, Hangzhou 310027, P. R. China
E-mail: {xin_yu; chenhf; wdzhao; xiel}@zju.edu.cn
Abstract—User experience becomes a most crucial factor for the
application and promotion of a new technology. Traditional
quality of service (QoS) can only measure the objective quality of
services and networks, while the quality of experience (QoE),
which has become a hot topic in recent years, can reflect
subjective feelings more directly from users’ perspective. In this
paper, we investigate the QoE evaluation method of video
streaming service in 3G networks, and propose a no-reference
QoE prediction model of video streaming service based on the
gradient boosting machine. Our proposed QoE prediction model
considered comprehensive parameters from the network layer,
the application layer, decoded videos and the user equipment.
Simulation results show that the performance of our proposed
QoE prediction model outperforms the G.1070 model, in terms of
accurate predicted mean opinion score, small root mean squared
error, and low time-consuming.
Keywords-Quality of service (QoS); Quality of experience (QoE);
Gradient boosting; Video quality assessment
I. INTRODUCTION
With the development of 3G networks and the emergence
of intelligent terminals, more and more people enjoy the video
streaming service through 3G networks and portable terminals.
Since the video streaming service demands a higher resource
than traditional services, such as voice and web browsing, how
to guarantee the user experience is a challenge issue to the
network operator. Quality of experience (QoE), defined as “the
overall acceptability of an application or service, as perceived
by the end-user” in [1], not only includes the quality of service
(QoS), but also considers the capability of the user equipment
(UE) as well as user’s expectation and context. Hence, QoE is a
comprehensive indicator to measure the performance of the
end-to-end systems, while QoS only considers the network
parameters, such as packet loss rate, delay, and so on, as shown
in Fig. 1.
Figure 1. QoE and QoS for video streaming service in 3G networks
Many methods have been proposed to evaluate the QoE
subjectively and objectively. In subjective quality assessment,
testers rate some degraded videos using Mean Opinion Score
(MOS) ranging from 1 to 5, where 1 means that the quality is
too poor to bear, and 5 means that the degradation is
imperceptible compared to the source videos [2]. Although
MOS can reflect the QoE most directly, it is time-consuming
and expends a large amount of manpower, which is not
practical and becoming a benchmark for the performance of
objective methods.
Objective quality assessment can be divided into three
categories, full-reference (FR) method, reduced-reference (RR)
method and no-reference (NR) method, according to whether
the source videos are used or not. Peak-signal-to-noise-ratio
(PSNR) and video quality measurement (VQM) [3] are the
common FR metrics. VQM models the human visual system in
the statistics theory, and its result is close to the perceived
MOS. To reduce computational complexity, RR models extract
only a part of features from the source videos. However, both
FR and RR methods need to use high quality reference videos,
which always cannot be accessible in reality. Thus, the NR
method is a feasible way to measure the QoE in real time by
only analyzing transmitted videos.
In [4], authors developed an NR video quality assessment
tool with blockiness and blur from compression domain. But
this assessment tool did not consider the impact of the network
transmission. In [5], authors used a decision tree to quantify the
QoE in real-life settings with multi-dimensional parameters.
However, with this method, the ratio of correct classification is
only 46.2%, and the performance is not good enough for
practicality. In [6], authors proposed a QoE space with
N reference points, and the prediction value equals to the
reference value with nearest Euclidean distance. But the time
complexity of this method is
2
()NΟ , which is not suitable for
a large scale data set.
In addition, an opinion model for video phone, only
considering the degradation introduced by network and
compression, is standardized as G.1070 [7] by ITU-T.
In this paper, we propose an end-to-end, no-reference and
real-time QoE prediction model for video streaming in 3G
networks. In our proposed model, the Gradient Boosting
This work was partly supported by Ministry of Industry and Information
Technology of P. R. China (No. 2012ZX030011035-004), National Natural
Science Foundation of China (No. 61071129, No. 61171087), and Science and
Technology Department of Zhejiang Province (No. 2011R10035).
978-1-61284-683-5/12/$31.00 ©2012 IEEE