See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/289760512

Survey on machine learning-based QoE-QoS correlation models

Article · January 2014

DOI: 10.1109/ComManTel.2014.6825604

CITATIONS READS

61 2,643

2 authors:

Aroussi Sana Abdelhamid Mellouk

Saad Dahlab University Université Paris-Est Créteil Val de Marne - Université Paris 12
6 PUBLICATIONS 113 CITATIONS 366 PUBLICATIONS 4,795 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Aroussi Sana on 19 March 2017.

The user has requested enhancement of the downloaded file.

 Survey on Machine Learning-based QoE-QoS
Correlation Models

Sana Aroussi Abdelhamid Mellouk

Computer Science Department, Sciences Faculty, Saad Image, Signal and Intelligent Systems Laboratory-LISSI &
Dahlab University, Blida Networks and Telecoms Dept-IUT C/V
National High School of Computer Science (ESI), Algiers University of Paris-Est Creteil (UPEC)
Algeria France
[email protected] [email protected]

Abstract— The machine learning provides a theoretical and an overview of the main QoE-QoS correlation models using
methodological framework to quantify the relationship between ML. Also, we propose a categorization of these models
user OoE (Quality of Experience) and network QoS (Quality of according to their learning type. For each category, we review
Service). This paper presents an overview of QoE-QoS the main existing works by citing deployed learning methods
correlation models based on machine learning techniques. and model parameters (QoE measurement, QoS parameters and
According to the learning type, we propose a categorization of service type). This survey will provide researchers with some
correlation models. For each category, we review the main of the latest trends and findings of this field.
existing works by citing deployed learning methods and model
parameters (QoE measurement, QoS parameters and service The rest of the paper is organized as follows. In Section 2,
type). Moreover, the survey will provide researchers with the we introduce the different types of ML. Then, we review the
latest trends and findings in this field. main existing QoE-QoS correlation models in Section 3 and 4.
Finally, in the discussion section, we draw conclusions.
Keywords-component; Machine Learning; Quality of
Experience; Quality of Service; Correlation model;
II. MACHINE LEARNING
I. INTRODUCTION Applied to the QoE-QoS relationship modelling, Machine
Learning (ML) methods use a set of observations reflecting the
The notion of Quality of Experience (QoE) has emerged network state and the user’s perception, in order to extract
during the last decade with the intention of involving the inference rules to predict automatically the QoE value. To deal
subjective perception of end-users into the evaluation of with this modelling problem, it is necessary to select the
telecommunication services quality. As is known to all, the appropriate learning type. In the following, we define the main
end-users dictate the success or failure of any service. So, learning types focusing on the ones that suit the QoE-QoS
services providers should satisfy their customers’ needs by relationship modelling.
offering high-performance services in order to prevent user
churn and save revenue [1]. The QoE determines the customer A. Deductive versus Inductive learning
satisfaction degree of a service.
The inference rules construction can be carried out in a
On the other hand, as telecommunication services rely on deductive or inductive manner. In deductive learning, the
interconnected network, service performance and thus particular rules are deducted from the general rules. In contrast,
customer satisfaction depend on network performance. Over with inductive learning, the general rules are drawn from the
the years, network operators have been assessing network particular cases observation. In the current paradigm, ML is
performance based on the Quality of Service (QoS) parameters considered to be the induction science [3]. The purpose of
such as throughput, delay, jitter, or loss rate. Consequently, it is learning methods is to find patterns from a collection of
of utmost importance to understand the relationships between specific observations so as to make predictions about the QoE
user-oriented QoE and network-oriented QoS parameters. This for upcoming events.
QoE-QoS relationship is currently subject of a huge increasing
interest both from academic and industrial perspectives. B. Supervised, Semi-supervised versus Unsupervised
At the crossroads of many disciplines such as statistics and Learning
artificial intelligence, Machine Learning (ML) provides a The goal of ML is to model an unknown target concept
theoretical and methodological framework to quantify the from observations. Depending on the nature of these
QoE-QoS relationship. Indeed, the ML includes a set of observations, three types of learning can be identified [4] [5]:
methods to build a reality model from data, either by
improving an existing model (partial or less general) or by • The supervised learning where the observations are
completely creating a new model [2]. In this paper, we present given in the form of input-output pairs. The purpose is

978-1-4799-2903-0/14/$31.00 ©2014 IEEE

 to learn a function explaining the relationship between depending on the QoE values type, offline batch models can be
the inputs and outputs. divided into two groups. The first group uses the regression
analysis to approximate the QoE as a continuous function of
• The unsupervised learning where the observations QoS parameters like in [8- 24].The second group uses
present only the input values. The purpose is to find classification methods to predict the QoE class [25-29]. In the
similarities between these values and to group similar following sections, we briefly describe these models.
data into clusters.
• The semi-supervised learning where the observations IV. OFFLINE BATCH QOE-QOS CORRELATION MODELS
are in the form of input-output pairs but the outputs
These prediction models are based on offline learning
values are not known in a large amount of
where a predefined data sample should be available during the
observations. The purpose is to use these latter to
batch training phase in order to find a deterministic relationship
improve the concept modelling in supervised learning.
between the QoS parameters (inputs) and the QoE (result).
Among these three types of learning, supervised and even Most models attempt to define a mathematical form of this
semi-supervised learning best fit the QoE-QoS correlation relationship using regression analysis.
modelling. Furthermore, if the QoE is expressed on a discrete
scale, it is a classification problem, otherwise, when it is A. Models using Regression Analysis
expressed on a continuous scale, it is the regression problem. Based on the regression analysis simplicity, several QoE-
QoS correlation models have been proposed. We differentiate
C. Offline versus Online Learning between those using Least Squares Regression (LSR) [8-20]
In offline learning, the model is first trained on a set of and those using Regression Neural Networks (RNN) [21-24].
collected data, then, deployed on the new data. However, the
online learning doesn’t separate the model training and its 1) Models using LSR:
deployment. It continues to learn as soon as new data are In [8], the authors analyzed the relationship between QoE,
available in order to improve its performance. Thus, the online expressed in Opinion Score (OS), and QoS parameters ,such
model can be adjusted very precisely to the end user [6]. as loss ratio, download time and throughput, in web surfing
service. To do this, different relationship forms (linear,
D. Batch versus Incremental Learning logarithmic, exponential and power) were analyzed and
evaluated through the correlation coefficient. The results show
The model training can be done in two ways [6]: batch or
that there is a linear relationship between QoE and loss ratio,
incremental. In batch learning, data are predefined and
available a priori. However, in incremental learning, data are which is exponential between QoE and download time and
introduced progressively and the model is able to change and logarithmic between QoE and throughput.
adjust its settings after observation of each data in order to The authors in [9], studied the impact of packet loss rate,
enrich their previously acquired learning from old data. The packet reorder rate and video bit rate on the QoE, measured in
QoE-QoS correlation models are generally designed to be used terms of PVQ (Perceived Video Quality) and PSNR (Peak
in real time and in dynamic environments. Thus, an Signal to Noise Ratio), in video streaming services.
incremental learning seems to be a promising solution. Indeed, Simulation results suggest (i) an exponential relationship
it can easily be made online, enabling to use the model during between QoE and QoS parameters namely packet loss and
it learning and to improve it throughout its use. packet reorder, and (ii) two relationship forms (logarithmic
and exponential) between video bit rate and QoE measurement
E. Passive versus Active Learning (PVQ and PSNR respectively).
The active learning enables the model to build the data set The QoE-QoS correlation models proposed in [8], [9] are
during its training by interacting with a human expert [7]. specific according to service types, and can’t be generalized
Learning starts with few data, then the model selects the most for any service type. However, there are three groups of
"relevant" data. Unlike passive learning where the collected solutions attempting to define a general relationship between
data are independent and identically distributed (i.i.d), active QoE and QoS, regardless of network service type (Table 1).
learning can accelerate the training phase considering the most
relevant data first. Therefore, such learning is useful when data
are too abundant and we must be able to lead the model
towards relevant data or on the contrary when the data are
costly to obtain.

III. QOE-QOS CORRELATION MODELS USING ML:

CATEGORIZATION
From the literature, several QoE prediction models using
ML techniques have been proposed [8-31]. We find that all
these models use an inductive supervised learning and can be
classified into two categories (Fig. 1): offline batch models [8-
29] and online incremental models [30] [31]. Furthermore, Figure 1. Categorization of QoE-QoS Correlation Models.
 TABLE I. GENERAL RELATIONSHIPS BETWEEN QOE AND QOS authors use the least squares method for linear multiple
Model Name Form Relation regression to define the relationship between QoE and
Reichl et al. Weber- multiple QoS parameters.
Logarithmic
[10] [11] Fechner Law In the framework of video streaming management with
Fielder et al. IQX heterogeneous access technologies, [19] proposed a specific
Exponential
[12-15] Hypothesis
Stevens’
function measuring the QoE (MOS) from a set of QoS
Khorsandroo
et al. [16-18] Power Law Power parameters (packet loss rate, frame rate, bandwidth, round trip
time and jitter). However, we have proposed in a previous
work [20] a global correlation model between QoE and several
The works of Reichel et al. [10] [11] present an example of the
QoS parameters based on IQX hypothesis. Indeed, our model
first solution. The authors claim that the QoE-QoS relationship
can be seen as an extension of the IQX hypothesis where QoS
is of logarithmic nature based on the psychological Weber-
is modeled as a weighted sum of several parameters. It has
Fechner Law (WFL). Their findings are validated
been validated for Video on Demand (VoD) services where
experimentally, where speech quality, measured by Mean
the video quality, expressed in Continuous Scale Opinion
Score Opinion (MOS) is considered as a logarithmic function
Score (OSCS), is considered as an exponential function of loss
of bit rate or loss rate in Voice over IP (VoIP) services [11].
ratio and delay.
The second solution is based on the IQX hypothesis
In summary, Table II lists all the works using LSR.
(Interdependency between QoE and QoS is eXponential) [12-
15]. In these articles, the authors assume that the QoE 2) Models using RNN: LSR methods are often insufficient
fluctuates according to its current level. This assumption leads to address the nonlinear multiple regression problem.
to a differential equation whose resolution gives an Therefore, researchers have turned to Artificial Neural
exponential relationship between QoE and QoS. The IQX Networks (ANN) which are good candidates for modelling the
hypothesis is validated for VoIP services where QoE is nonlinear relationship between QoE and multiple QoS
expressed in terms of MOS as functions of loss or reordering parameters [32]. Table III summarizes the main prediction
ratio. In addition, the authors show that the exponential model models based on ANN. We find that these models use
provides approximations with better quality than the multilayer feedforward neural network consisting of a layer of
logarithmic model proposed in [8] for web surfing service. input neurons, one or more layers of hidden neurons, and one
The third solution is proposed in the works of Khorsandroo et layer of output neurons. The number of input neurons
al. [16-18]. In [16], the results indicated that it is possible to represents the number of selected QoS parameters. There is a
consider a relation in the form of a power function to explain single output neuron corresponding to the QoE measurement.
the possible relationship between QoE (MOS) and QoS The learning algorithm used is gradient descent
(packet loss) in video streaming services. Then, in [17], the backpropagation. Note here that [22] ([23] respectively)
authors introduced the psychophysics Stevens’ Power Law in generates several RNN according to the QoE metrics and the
order to show QoE-QoS correlation in the form of a power video content type (the router queue discipline, the codec type
function. A theoretical and empirical comparison was made and resolution respectively).
with the Weber-Fechner law proposed in [11]. The results
showed that logarithmic form can serve better than a power B. Models using Classification Methods
form solution, but cannot be a dominant solution. In this category, few studies have been conducted. Agboma
Recently, in [18], Khorsandroo et al., have completed their and Liotta [25] [26] have used Linear Discriminant Analysis to
studies with IQX hypothesis. According to how they define a predict the QoE class (acceptable or unacceptable) from two
general relationship between QoE and QoS, these relations QoS parameters: video bit rate and frame rate. Depending on
(logarithmic, exponential and power) are classified as the terminal and video content type, this work has generated
stimulus-centric and perception-centric relations. The first several prediction models with a precision between 75-85%.
class (psychophysical laws) considers QoE as a function of To enhance this accuracy rate, Menkovski et al. used in [27]
QoS parameters changes, while the latter (IQX hypothesis) [28] other learning methods like Support Vector Machines
states that QoE changes in accordance with the current level of (SVM) and Decision Trees (DT). In fact, these two methods
the user perception. These two classes were analyzed and have generated models with high accuracy (over 90%) where,
compared theoretically and practically. The results led the for each terminal type (Mobile, PDA, Laptop), the QoE class
authors to introduce a new definition of QoE-QoS relationship (acceptable or unacceptable) is estimated from four video
which results in a complete differential form of function. This parameters: bit rate, frame rate, spatial and temporal
function can explain how QoE changes in two different information. The SVM and DT methods were also used in [29]
aspects (stimulus-centric and perception-centric) at a same and compared with other classification methods, namely Naive
time. Bayes (NB), k-Nearest Neighbours (k-NN), Random Forest
In the above mentioned works, we note that the authors use (RF) and ANN. The results show that the RF and DT methods
the least squares method for a simple regression. In other are most appropriate for estimating the QoE class (five classes
words, they study the impact of a single QoS parameter on the according to the MOS score) from different parameters (delay,
QoE. However, there are other works that take into account jitter, packet loss, video types, users’ profile and terminal
several QoS parameters like [19] [20]. In these papers, the characteristics).
 V. ONLINE INCREMENTAL QOE-QOS CORRELATION literature of QoE-QoS correlation models using ML. All these
MODELS models use inductive supervised learning techniques and most
of them are built in offline batch manner using regression
To the best of our knowledge, only the works of Menkovski
analysis. Moreover, some models focus on QoE-QoS
et al. in [30] [31] propose online QoE prediction model. This
dependency in a specific service, while others have tried to
model can estimate the QoE based on given QoS metrics from
devise a general solution regardless of service type. We think
continuous real-time user feedback and does not require a
that a general model cannot be a dominant solution to define
priori execution of subjective studies. To do this, the authors
the QoE-QoS relationship because the QoS parameters and
have tested several variants of online learning algorithms based
their individual and collective impact on the QoE depend on
on Hoeffding Option Trees, Naïve Bayes and OzaBagging
the service type.
ensemble. For the needs of their application (mobile
multimedia streaming services), they found, in [30], that the On the other hand, the QoE prediction models enable the
most suitable algorithm is the OzaBagging ensemble with network and service providers to monitor and react upon
Hoeffding Option Tree Naïve Bayes Adaptive. quality problems, at best before the customer perceives them.
Thus, they should estimate the user perception based on
VI. DISCUSSION AND CONCLUSION continuous real-time QoE-QoS feedback from dynamic
networks. Online incremental techniques seem to be a
Quantifying the QoE-QoS relationship is an extremely promising solution. We think that the existing offline batch
challenging task and the ultimate judge of telecommunication models need to be adapted using online incremental learning
service quality. A number of researches have been selected and algorithms like in [30] [31].
analyzed in this paper in order to obtain a global view of the

TABLE II. A SUMMARY OF MODELS USING LSR

Simulation results
Relation Regression
Model Form QoS Parameters QoE Service Type
Type Type
measurement
Linear Loss ratio

[8] Specific Simple Exponential Download time OS Web surfing

Logarithmic Throughput

Packet loss or Packet reorder PVQ, PSNR

Exponential
[9] Specific Simple PVQ video Streaming
Video bit rate
Logarithmic PSNR

[11] General Simple Logarithmic Bit rate or Loss rate MOS VoIP

[13][14] Loss ratio or Reordering ratio MOS VoIP

General Simple Exponential
[15]
[16] General Simple Power Packet loss MOS Video streaming

Packet loss rate, Frame rate, Bandwidth, Round trip time and MOS Video streaming
[19] Specific Multiple Linear
Jitter
[20] General Multiple Exponential Loss ratio and delay OSCS VoD

TABLE III. A SUMMARY OF MODELS USING RNN

Models [21] [22] [23] [24]
ANN Architecture Multilayer Feedforward Neural Network
Learning Algorithm Gradient Descent Backpropagation
QoS Parameters Delay, Jitter, Loss ratio, Bandwidth, Delay, Jitter and Jitter, Loss ratio and Bit Jitter, Loss ratio, Bandwidth, Frame
Burst, Congestion period and Disorder Loss ratio rate. rate, Bit rate Sender, video content
packets. type and screen type.
QoE Measurement DMOS MOS, PSNR, MOS MOS
SSIM, VQM
Service Type Video streaming Video streaming IPTV Video streaming
 Streaming Case Study,” International Conference on Advances in
VII. REFERENCES Computing, Communications and Informatics (ICACCI), Chennai, T
Nadu, India, pp. 446–452, 2012.
[1] Abdelhamid Mellouk, Said Hoceini, and Hai Anh Tran, Quality of
[18] S. Khorsandroo, R. M. Noor, and S. Khorsandroo, “A Generic
Experience for Multimedia, Application to Content Delivery Network
Quantitative Relationship to Assess Interdependency of QoE and
Architecture, WILEY. 2013.
QoS,” KSII Trans. INTERNET Inf. Syst., vol. 7, no. 2, pp. 327–346,
[2] A. Cornuéjols and L. Miclet, Apprentissage artificiel : concepts et
2013.
algorithmes, EYROLLES. 2003.
[19] M. Elkotob, D. Grandlund, K. Andersson, and C. Ahlund,
[3] A. Cornuéjols, “Apprentissage et Circulation d’Information,” HDR,
“Multimedia QoE optimized management using prediction and
Université Paris–Sud, France, 2005.
statistical learning,” IEEE Conference on Local Computer Networks,
[4] H.-M. Suchier, “Nouvelles Contributions du Boosting en
Denver, Colorado, USA, pp. 324–327, 2010.
Apprentissage Automatique,” Thèse de doctorat, Université Jean
[20] S. Aroussi, T. Bouabana-Tebibel, and A. Mellouk, “Empirical QoE-
Monnet de Saint-Etienne, France, 2006.
QoS Correlation Model based on Multiple Parameters for VoD
[5] P. DALLAIRE, “Apprentissage par Renforcement Bayesien de
flows,” Global Communications Conference (GLOBECOM),
processus décisionnel de Markov partiellement observable : une
Californie, USA, pp. 1963 – 1968, 2012.
approche basée sur les processus Gaussiens.,” Université Laval,
[21] H. Du, C. Guo, Y. Liu, and Y. Liu, “Research on Relationships
QUEBEC, 2010.
between QoE and QoS based on BP Neural Network,” IEEE
[6] M. Bouillon, “Apprentissage incrémental et décrémental:
International Conference on Network Infrastructure and Digital
Classification avec un système d’inférence floue évolutif appliquée à
Content (IC-NIDC), Beijing, China, pp. 312–315, 2009.
la reconnaissance de gestes manuscrits,” Thèse de master, Institut
[22] V. A. Machado, C. N. Silva, R. S. Oliveira, A. M. Melo, M. Silva, C.
National des Sciences Appliquées (INSA) de Rennes, France, 2012.
R. L. Francês, J. C. W. A. Costa, N. L. Vijaykumar, and C. M. Hirata,
[7] A. Bondu and V. Lemaire, “État de l’art sur les méthodes statistiques
“A New Proposal to Provide Estimation of QoS and QoE over
d’apprentissage actif,” Rev. Nouv. Technol. L’Information RNTI, p.
WiMAX Networks: An approach based on computational intelligence
19, 2007.
and discrete-event simulation,” IEEE Latin-American Conference on
[8] J. Shaikh, M. Fiedler, and D. Collange, “Quality of Experience from
Communications (LATINCOM), Belem do Para, Brésil, 2011.
user and network perspectives,” Ann. Telecommun., vol. 65, pp. 47–
[23] P. Calyam, P. Chandrasekaran, G. Trueb, N. Howes, R. Ramnath, D.
57, 2010.
Yu, Y. Liu, L. Xiong, and D. Yang, “Multi-Resolution Multimedia
[9] K. ur R. Laghari, O. Issa, F. Speranza, and T. H. Falk, “Quality-of-
QoE Models for IPTV Applications,” International Journal of Digital
Experience Perception for Video Streaming Services Preliminary
Multimedia Broadcasting, p. 14, 2012.
Subjective and Objective Results,” APSIPA Annual Summit and
[24] B. Lahouari, “Mise en œuvre d’un modèle de corrélation QoS-QoE,”
Conference (ASC), Hollywood, California, USA, p. 9, 2011.
Université de Bretagne Occidentale, France, RAPPORT DE STAGE,
[10] P. Reichl, B. Tuffin, and R. Schatz, “Economics of Logarithmic
2012.
Quality-of-Experience in Communication Networks,” Conference of
[25] F. Agboma and A. Liotta, “QoE-aware QoS Management,”
Telecommunication, Media and Internet Techno-Economics (CTTE),
International Conference on Advances in Mobile Computing &
Ghent, Belgique, pp. 1–8, 2010.
Multimedia (MoMM), Linz, Austria, Autriche, pp. 111–116, 2008.
[11] P. Reichl, S. Egger, R. Schatz, and A. D’Alconzo, “The Logarithmic
[26] F. Agboma and A. Liotta, “Quality of experience management in
Nature of QoE and the Role of the Weber-Fechner Law in QoE
mobile content delivery systems,” Telecommun. Syst., vol. 49, pp. 85–
Assessment,” IEEE International Conference Communications (ICC),
98, 2012.
Cape Town, South Afric, pp. 1–5, 2010.
[27] V. Menkovski, A. Oredope, A. Liotta, and A. C. Sánchez, “Predicting
[12] T. Hoßfeld, P. Tran-Gia, and M. Fiedler, “Quantification of Quality of
Quality of Experience in Multimedia Streaming,” International
Experience for Edge-Based Applications,” LNCS 4516, pp. 361–373,
Conference on Advances in Mobile Computing & Multimedia
2007.
(MoMM), Kuala Lumpur, Malaysia, pp. 52–59, 2009.
[13] T. Hoßfeld, D. Hock, P. Tran-Gia, K. Tutschku, and M. Fiedler,
[28] V. Menkovski, A. Oredope, A. Liotta, and A. C. Sánchez, “Optimized
“Testing the IQX Hypothesis for Exponential Interdependency
online learning for QoE prediction,” Benelux Conference on Arti_cial
between QoS and QoE for Voice Codecs iLBC and G.711,”
Intelligence, Eindhoven, The Netherlands, pp. 169–176, 2009.
University of Würzburg, Research Report Series 442, 2008.
[29] M. S. Mushtaq, B. Augustin, and A. Mellouk, “Empirical Study based
[14] M. Fiedler, T. Hossfeld, and P. Tran-Gia, “A Generic Quantitative
on Machine Learning Approach to Assess the QoS/QoE Correlation,”
Relationship between Quality of Experience and Quality of Service,”
European Conference on Networks and Optical Communications
IEEE Netw., pp. 36–41, 2010.
(NOC), Vilanova i la Geltru, Espagne, pp. 1–7, 2012.
[15] M. Fiedler and T. Hossfeld, “Quality of Experience-related
[30] V. Menkovski, G. Exarchakos, and A. Liotta, “Online Learning for
differential equations and provisioning-delivery hysteresis,” ITC
Quality of Experience Management,” Annual Machine Learning
Specialist Seminar on Multimedia Applications - Traffic, Performance
Conference of Belgium and The Netherlands, Leuven, Belgium, p. 6,
and QoE, Myazaki, Japan, pp. 19–24, 2010.
2010.
[16] S. Khorsandroo, R. M. Noor, and S. Khorsandroo, “A Generic
[31] V. Menkovski, G. Exarchakos, and A. Liotta, “ONLINE QoE
Quantitative Relationship Between Quality of Experience and Packet
PREDICTION,” International Workshop on Quality of Multimedia
Loss in Video Streaming Services,” International Conference on
Experience (QoMEX), Trondheim, Norvège, pp. 118–123, 2010.
Ubiquitous and Future Networks (ICUFN), Phuket, Thailand, pp.
[32] G. DREYFUS, “LES RÉSEAUX DE NEURONES,” Mécanique Ind.
352–356, 2012.
Matér., vol. 51, p. 19, 1998.
[17] S. Khorsandroo, R. M. Noor, and S. Khorsandroo, “The Role of
Psychophysics Laws in Quality of Experience Assessment – A Video

View publication stats