Institute of Computer ScienceChair of Communication Networks

Prof. Dr.-Ing. P. Tran-Gia

The Memory Effect and Its Implications on Web QoE ModelingImplications on Web QoE Modeling

Tobias Hoßfeld 1,2, Sebastian Biedermann1, Raimund Schatz2, Alexander Egger2, Sebastian Egger2, Markus Fiedler3

Trend towards Quality of Experience

� Increasing competition among Telco‘s and ISPs, among application and service providers, among cloud providers� Keep customers happy, attract new customers � Quality as key differentiator, but only as experienced by end user

� Shift from Quality of service (QoS) to Quality of Experience (QoE)� QoS: packet loss, delay, jitter, …

Tobias Hossfeld 2

� QoE: subjective experience/satisfaction of users of a service� Example: VoIP user interested in speech quality

web user interested in short page load times

� What are relevant QoE influence factors ? � How to integrate key influence factors in appropriate QoE models ?

2

3

4

5

6

7

8

traf

fic in

crea

se (

norm

aliz

ed to

201

0)

Internet videoOnline gamingWeb, email, and dataFile sharingVoIP

QoE Model for Web Browsing

� Importance of web traffic is increasing� Related work on QoE mainly

covers multimedia applications� Scenario: User downloads several

web pages within a session� But: Random test sequences

according to standardization

Mainly web traffic

Tobias Hossfeld 3

2010 2011 2012 2013 2014 20151

2

year

traf

fic in

crea

se (

norm

aliz

ed to

201

0)

according to standardization� Lack of web QoE models including

quality changes over timeSource: Cisco Visual Networking Index: Forecast and

Methodology, 2010-2015

Contribution:� Subjective user study on web browsing with quality changes� Identification of memory effect as relevant QoE influence factors� Integration of key influence factors in appropriate QoE models

Agenda

� Conducted Subjective User Study � QoE Influence Factors, Design of Study� Implementation and Measurement Setup

� Statistical Analysis of User Ratings� Page Load Times, Memory Effect � Key QoE Influence Factor via Support Vector Machines

Tobias Hossfeld 4

� Key QoE Influence Factor via Support Vector Machines

� Implications on QoE Models � Iterative Regression Model� Hidden Memory Markov Model

� Conclusions and Outlook

QoE Influence Factors for Web Browsing

QoE

Context

User Expectations

Memory Effects

Usage History

Task/purpose e.g. time killing, information retrieval

Social and cultural background

�Quality

change per

page

�Online and

crowd-

sourcing tests

Tobias Hossfeld 5

System

Content Type of web site, e.g. YouTube, Google, Facebook

Transmission network, e.g. bandwidth, delay, jitter, loss

Client Device

Web browser

Implementation of web site

� Simple

Photo Page

� Emulate

page load

time

Methodology: Subjective User Studies

� Subjective user rests required due to lack of existing studies

Picture is downloaded with predefined time.

Test run over several web

Tobias Hossfeld 6

� Laboratory tests to get first insights, delay via traffic shaper� Online tests to reach more users, local applet with def. page load

times

User rating on 5-point ACR scale

Test run over several web pages (with changing network conditions)

Quantifying Quality of an Experience

Excellent

Good

5

4

Imperceptible

Perceptible

MOS Quality Impairment

Mean Opinion Score (MOS): numerical indication of t he perceivedquality of received media after compression and/or transmission

Excellent!

Bad!Poor!

∅∅∅∅

Fair = 3

Tobias Hossfeld 7

Good

Fair

Poor

Bad

4

3

2

1

Perceptible

Slightly annoying

Annoying

Very annoying

Fair!Good!∅∅∅∅

Agenda

� Conducted Subjective User Study � QoE Influence Factors, Design of Study� Implementation and Measurement Setup

� Statistical Analysis of User Ratings� Page Load Times, Memory Effect � Key QoE Influence Factor via Support Vector Machines

Tobias Hossfeld 8

� Key QoE Influence Factor via Support Vector Machines

� Implications on QoE Models � Iterative Regression Model� Hidden Memory Markov Model

� Conclusions and Outlook

Impact of Page Load Times on QoE

� Models from literature for mapping current QoS to QoE� IQX hypothesis when using network parameters like bandwidth

[Hoßfeld, Fiedler, Tran-Gia, ITC 2007]

QoE(x) = a exp(-b x) + c

� Weber-Fechner law from psychophysics when using page load time as QoS parameter [Reichl et al., ICC 2010]

QoE(x) = a ln(b x)

Tobias Hossfeld 9

� But: strong deviations frommodel observable� Only current QoS is

considered� Temporal effects have to

be included

The Memory Effect

Same QoS but different QoE

Good QoSbefore

Bad QoSbefore

Tobias Hossfeld 10

� Web pages with same page load time have different QoE, depending on previous QoS conditions� Exponential decays/increase after quality changes

before

Is Memory Effect a Relevant QoE Influence Factor?

� Investigation with correlation analysis and machine learning� Support vector machine decide two-class problems � user ratings

are separated into ‘good quality’ and ‘bad quality’� Implementation of SMO (Sequential Minimal Optimization) in

WEKA machine learning software for analysis

� QoE from previous web site has relevant

Tobias Hossfeld 11

web site has relevant impact on QoE� Memory effect as

dominant as technicalinfluences (PLT)� Only last QoE has to

be considered

Agenda

� Conducted Subjective User Study � QoE Influence Factors, Design of Study� Implementation and Measurement Setup

� Statistical Analysis of User Ratings� Page Load Times, Memory Effect � Key QoE Influence Factor via Support Vector Machines

Tobias Hossfeld 12

� Key QoE Influence Factor via Support Vector Machines

� Implications on QoE Models � Iterative Regression Model� Hidden Memory Markov Model

� Conclusions and Outlook

Memory has to be included in QoE Models

� Support Vector Machines� Consider previous experiences as own variables� Weighting factors indicate ‘importance’ of variables

� Exponential iterative regressions� Weber-Fechner law yields f(PLT)� Considering previous QoE via iterations

Tobias Hossfeld 13

� Considering previous QoE via iterations

� Markov models� are memoryless,� but can include memory by appropriate state space

Hidden Memory Markov Model

� Describe the QoE with a Hidden Markov Model (HMMM)� Page load time as hidden state� QoE i.e. user ratings as emission

� Sequence of web pages withpage load times xi is extendedto series of pairs (xi,xi-1)Page load times are discretized

Hi, Hi-1

transition prob.

emission prob.

Tobias Hossfeld 14

� Page load times are discretized

� 2D-State space of HMMM is (Hi,Hi-1)� Transition and emission probabilities

derived from user studies

Conclusions

� Time dynamics of human perception for web QoE analyzed� Designed and conducted subjective user study on web browsing

� Identification of memory effect as relevant QoE influence factors� Integration of key influence factors in appropriate QoE models

� Support vector machines with additional ‘past’ variables� Weber-Fechner law with iterative exponential regressions

Tobias Hossfeld 15

� Weber-Fechner law with iterative exponential regressions� Hidden Markov model by increasing state space

� Consequences � QoE model available for performance evaluation,

measurement studies, subjective user surveys � In case of unpredictable QoS: avoid memory effects (e.g. for

QoE based traffic management, development of apps, etc.)� In general: Interdisciplinary research required

Institute of Computer ScienceChair of Communication Networks

Prof. Dr.-Ing. P. Tran-Gia

?

??

?

?

?

?

??

??

?

?? ?

?

?

?

?

?

??

?

?

??

??

?

?

?

?

??

?

??

??

?

??

?

??

?

?

?

?

? ?

?

??

??

?

?

?

?

?

?

??

?

?

?

??

? ??

?

?

??

? ?

??

?

?

?

??

? ??? ?

??

?

?

?

?

?

?

?

? Questions?

?

?

?

?

? ??

??

??

??

? ?

??

?

?

??

?

?

?

?

? ?

?

? ??

?

?

??

?

?

?

??

?

?? ??

?

??

??

?

?

??

?

? ??

?

?

?

?

? ?

?

?

?

?

??

?

??

? ?

??

?

??

? ?

?

?

?

??

? ??

??

?

Local Tests in Laboratory Environment

Tobias Hossfeld 17

Online Tests with “Net-Sim -Applet”

Tobias Hossfeld 18

Statistics about the Conducted Experiments

Exp. Id #test users

X-point scale

#web pages

#changes of PLT

MinPLT (ms)

Mean PLT (ms)

Max PLT (ms)

0 29 3 93 21 348 2594 8184

1 12 3 40 2 240 420 720

2 72 5 40 7 240 660 1200

Tobias Hossfeld 19

3 30 3 25 4 240 336 480

4 26 5 40 9 240 600 960

5 15 5 25 4 240 528 720

Local Test Run (Exp. 0)

� Page download times measured via TCPDump / HTTPFox

6

7

8

9

time

(s)

local test run

Tobias Hossfeld 20

0 20 40 60 80 1000

1

2

3

4

5

web page

page

dow

nloa

dtim

e (s

)

0 10 20 30 400

0.51

1.5

exp. 1

0 10 20 30 400

0.51

1.5

exp. 2

1.5

page

dow

nloa

d tim

e (s

)Online Test Runs

� Test memory effect

� Test several effects

Tobias Hossfeld 21

0 5 10 15 20 250

0.51

1.5

exp. 3

0 10 20 30 400

0.51

1.5

web page

exp. 4

page

dow

nloa

d tim

e (s

)

� Test. exp. regressions

� Test several effects

Agenda

� Existing QoE Models for Web Traffic

� Measurement Settings

� QoE Characteristics

Tobias Hossfeld 22

� QoE Characteristics

� Identification of User Groups

� QoE Models

Existing QoE Models for Web Traffic

� Fidel Liberal, Armando Ferro, Jose Oscar Fajardo: “PQoS based model for assessing significance of providers statistically” (2005) � MOS = 6 - log2(t), t: total download time [s]

� R.E. Kooij, R.D. van der Mei, R. Yang: “TCP and web browsing performance in case of bi-directional packet loss” (2009)� MOS = 4.75 - 0.81log(t), t: total download time [s]

� ITU-T Recommendation G.1030: “Estimating end-to-end performance in IP

⋅ ⋅

Tobias Hossfeld 23

� ITU-T Recommendation G.1030: “Estimating end-to-end performance in IP networks for data applications” (2005)� MOS = 4.298⋅exp(−0.347⋅t)+1.390, t: weighted session time [s]

� J. Shaikh, M. Fiedler, D. Collange, “Quality of Experience from user and network perspectives” (2010)� MOS = 1.15 + 1.50 ln(R), R: delivery bandwidth [Mbps]� MOS = 5.50 exp(−0.2L) , L: packet loss ratio [%]

� What’s missing? Psychological and temporal effects!

QoE Related to Page Download Time

3

3.5

4

4.5

5M

OS

LiFeFa05KoMeYa09G.1030

Tobias Hossfeld 24

0 5 10 151

1.5

2

2.5

3

total download time [s]

MO

S

QoE Depending on Bandwidth and Loss

10

20de

liver

y ba

ndw

idth

[Mbp

s]

ShFiCo100

5

pack

et lo

ss r

atio

[%]

Tobias Hossfeld 25

1 2 3 4 50

10

deliv

ery

band

wid

th [M

bps]

MOS1 2 3 4 5

5

10

pack

et lo

ss r

atio

[%]

Overview on Test Runs

Tobias Hossfeld 26

Exponential Regression

� Exponential decay / increase (similar to Raake, A. (2006a) Short-and long-term packet loss behavior: towards speech qualityprediction for arbitrary loss distributions)

Tobias Hossfeld 27

Simple Cluster Analysis

� For each user, the average rating and the coefficient of variation of the user ratings is calculated

� Cluster analysis with k-means algorithm (Matlab, RapidMiner)� As input parameter,

only these two 0.5

0.6

0.7

0.8

coef

ficie

nt o

f var

ianc

e

Cluster 1Cluster 2Cluster 3

Tobias Hossfeld 28

only these two parameters are used� Simple approach is

already sufficient to detect the different clusters

1.5 2 2.5 3 3.5 4 4.5 50

0.1

0.2

0.3

0.4

average rating of each user

coef

ficie

nt o

f var

ianc

e

QoE Ratings for Different Clusters

3

4

5

aver

age

user

rat

ing

of th

ree

grou

ps, d

ownl

oad

time

Cluster 1:

bored users

Cluster 3:

Tobias Hossfeld 29

0 5 10 15 20 25 30 35 40

1

2

sequence of web pages

aver

age

user

rat

ing

of th

ree

grou

ps, d

ownl

oad

time

Cluster 3:

normal users

Cluster 2:

hectic users

Implicit QoE Model: Hidden Markov Model

� Describe the QoE with a Hidden Markov Model (HMM)� Download time as hidden state� QoE / user ratings as emission

� State transition matrix describessystem dynamics (in terms of QoS)

Tobias Hossfeld 30

� Emission probabilities for perceptioncategories (MOS 1, …, 5) accordingto actual state and user group

� One-dimensional HMM fails, since memory effect is not taken into account

2D Hidden Memory Markov Model

� Enhance the HMM by one dimension wrt. memory effect� State of the system is a tupel

(actual download time dt_i, previous download time dt_i-1)� QoE / user ratings as emission

Tobias Hossfeld 31

Notes about the 2D -HMMM

� Relevant outcome of subjective tests are emission probabilities (for given, i.e. tested, network tupels)

� Underlying network model (i.e. hidden states) can be changed� for a proper description of a system wrt. QoE, it is important to

describe it as 2D MMM (due to memory effect)� then emissions probabilities remain the same and can be

applied

Tobias Hossfeld 32

applied

� Problem is to get measurement data for all N2 states, when Ndownload times are observed

Discussion

� Discrete States of HMM� Weber’s law from psychophysics (1840): The just-noticeable

difference (∆R) of the change in a stimulus’s magnitude is proportional to the stimulus’s magnitude (R), rather than being an absolute value.

� ∆R / R = k and states of HMM are defined accordingly

� State of the system

Tobias Hossfeld 33

� State of the system� previous download time vs. average download time (using

exponential moving average and discretization of download times)

� Similar to Oliver Rose: A Memory Markov Chain for VBR traffic with strong positive correlations, ITC 16, Jun 1999.

Test Run 2: Complete User Survey

� Simulated QoS settings in terms of page download time are colored according to MOS� Complete CDF gives overview on actual user experience, not on

average users (MOS1+MOS5 ~= MOS3+MOS3)

1 MOS1 MOS2 MOS3 MOS4 MOS5

Tobias Hossfeld 34

0 10 20 30 400

0.2

0.4

0.6

0.8

web page

ratio

of u

sers

Weber-Fechner Law

� Web traffic experiments@UniWue

3.5

4

4.5

5Q

oEWeber-Fechner-Law for web traffic

Tobias Hossfeld 35

102

103

1041

1.5

2

2.5

3

page load time [s]

QoE

exp.0, R2 = 0.968

exp.2, R2 = 0.939

exp.4, R2 = 0.854