PaperlayerView exact revision on arXiv

Source-linked AI summary

Survey on QoE\QoS Correlation Models For Multimedia Services

Mohammed Alreshoodi, John Woods

arXiv:1306.0221v1cs.MMcs.NI

TL;DR

The paper addresses the poorly understood quantitative relationship between QoS and subjective QoE, for which existing mappings are often only partial. It reviews QoE/QoS correlation models and optimisation approaches, concluding that optimised coefficients and quantitative analyses combining subjective and objective measures better reflect QoE complexity.

  • Problem

    QoS–QoE interactions remain poorly understood, with no standardized methodology that directly and quantitatively maps QoS to QoE.

  • Method

    The paper reviews QoE/QoS correlation models across multimedia services, examining measurement approaches, quality metrics, QoS parameters, and influencing factors.

  • Results

    Models using optimised coefficients and artificial optimisation techniques can find better-fitting QoE mappings by learning nonlinear QoS–QoE relationships.

  • Takeaways & Limitations

    Quantitative subjective measures enable combined statistical analysis with objective variables, while user-behaviour measures can provide objective indicators of commonly subjective experience.

  • Takeaways & Limitations

    Some reviewed studies use generic relationships involving a single QoS parameter per experiment and require network emulators to represent burst losses.

Abstract

from arXiv · show

This paper presents a brief review of some existing correlation models which attempt to map Quality of Service (QoS) to Quality of Experience (QoE) for multimedia services. The term QoS refers to deterministic network behaviour, so that data can be transported with a minimum of packet loss, delay and maximum bandwidth. QoE is a subjective measure that involves human dimensions; it ties together user perception, expectations, and experience of the application and network performance. The Holy Grail of subjective measurement is to predict it from the objective measurements; in other words predict QoE from a given set of QoS parameters or vice versa. Whilst there are many quality models for multimedia, most of them are only partial solutions to predicting QoE from a given QoS. This contribution analyses a number of previous attempts and optimisation techniquesthat can reliably compute the weighting coefficients for the QoS/QoE mapping.

1. INTRODUCTION

The paper distinguishes objective QoS from subjective QoE and reviews efforts to quantify their relationship for multimedia services. It emphasizes that user satisfaction depends on both network conditions and human factors.

  • QoS characterizes deterministic network behavior through packet loss, delay, and bandwidth, whereas QoE incorporates user perception and other human dimensions.
  • About 90% of users reportedly leave providers rather than complain when service quality is low, motivating continual QoE measurement and improvement.
  • Perceived quality is influenced by network reliability, content preparation, terminal performance, and subjective factors such as experience, interest, and expectation.
  • Existing methods vary by media type and measurement requirements, while the paper reviews QoE/QoS correlation approaches and their quantitative relationship.

2. BACKGROUND

The background presents QoS and QoE layers, assessment methodologies, model classifications, and mapping functions used to estimate subjective quality from objective measurements.

  • QoS and QoE Layers: Application-layer QoS includes frame rate, resolution, color, and codec type, while network-layer QoS includes delay, jitter, and packet loss.
  • QoS and QoE Layers: QoE generally decreases as QoS disturbance increases, with a three-zone mapping from unaffected appreciation to possible service abandonment.
  • QoE Measurement approaches: MOS is a standardized subjective score from 1 to 5, but subjective assessment is costly, time-consuming, non-real-time, and difficult to reproduce.
  • QoE Measurement approaches: Objective assessment produces quantitative video-quality measures and may use intrusive signal-based or non-intrusive network/application-parameter methods.
  • QoE Measurement approaches: User-perspective measurement approaches comprise Testing User-perceived QoS, Surveying Subjective QoE, and Modelling Media Quality.
  • Classification of Objective Quality Assessment Models: Objective quality models include parametric packet-layer, parametric planning, media-layer, bit-stream, and hybrid models with different input requirements and limitations.
  • Mapping Function: Mapping functions transform objective video quality into predicted MOS; linear fits suit uniformly scaled scores, whereas nonlinear functions address nonuniform quality scales.

3. QOS/QOE CORRELATION APPROACHES

The reviewed correlation approaches model QoE as a function of QoS, including the exponential IQX hypothesis. In the cited study, IQX approximations outperformed a logarithmic alternative but excluded time-varying impairments.

  • The paper surveys general QoE/QoS correlation models to provide a broader picture of multimedia-service modelling approaches.
  • IQX hypothesis: One study assumed a linear dependence of QoE variation on QoS disturbances and derived the exponential IQX hypothesis.
  • IQX hypothesis: The IQX hypothesis expresses QoE as QoE = α * exp (− β * QoS) + γ, with positive parameters α, β, and γ.
  • IQX hypothesis: Using MOS as QoE and packet loss, jitter, response, and download times as QoS criteria, the study tested IQX with VoIP and web browsing.
  • IQX hypothesis: The proposed IQX approximations were better than the logarithmic expression, but the model did not capture time-varying IP impairments.

3.2. VQM-based Mapping Model

This model maps multidimensional QoS to QoE through the Video Quality Metric (VQM), using packet loss, delay, and jitter under specified streaming-video conditions. The derived formula captures observed relationships but was judged insufficient as a causal mapping without more complex fitting.

  • The model represents QoE with VQM as a function of n-dimensional QoS parameters.Here, VQM is treated as an indicator of QoE.
  • The study modeled packet loss, delay, and jitter using emulation results and derived a QoS-to-VQM formula for streaming video.The formula was fitted with n = 2 for the final expression and verified using additional video samples.
  • Increasing packet loss increased the VQM score, indicating worse QoE, but the fitted equation was not considered the best fit or sufficient to prove causality.The authors argued that more complex curve-fitting algorithms are needed.

3.3. QoEModel usingStatistical Analysis method

The statistical model reduces subjective testing by using limited viewer judgments and discriminant analysis to predict QoE groups from quantitative video parameters. Its accuracy varied by terminal, while limited predictors and the statistical method constrained the approach.

  • Limited subjective testing used quality-order judgments and the method of limits to identify noticeable video-quality changes.Discriminant Analysis then predicted group memberships from quantitative variables.
  • The discriminant score is modeled from quantitative predictors using fitted coefficients for each variable.The predictors are represented for each case and group, with coefficients indexed by variable.
  • Accuracy was 76.9% for mobile phones, 86.6% for PDAs, and 83.9% for laptops.The model used bitrate and frame rate across three terminals and six video-content types.
  • The approach had limited accuracy because of the statistical method and did not specify Network Layer QoS implementation.The authors also suggested adding video-content and coding parameters to improve perceived-quality prediction and network utilization.

3.4. QoEModels based onMachine Learning methods

Machine-learning approaches extend statistical QoE prediction with decision trees, support vector machines, neural networks, and neuro-fuzzy or regression models. Reported accuracy was generally strong, but parameter selection and further subjective validation remained important constraints.

  • Decision Trees and Support Vector Machines extended earlier QoE prediction models, with reported accuracies generally above 90% under cross-validation.Decision Trees achieved 93.55%, 90.29%, and 95.46% across mobile phones, PDAs, and laptops; SVM achieved 88.59%, 89.38%, and 91.45%.
  • Random Forest performed slightly better than Naive Bayes, k-NN, Neural Networks, Decision Trees, and SVM in a comparative machine-learning study.
  • An ANN model used network QoS parameters as inputs and MOS, PSNR, SSIM, and VQM as outputs, learning weights from a video database.
  • ANFIS and nonlinear regression models predicted MOS from physical- and application-layer parameters and produced good accuracy on an unseen dataset.The authors emphasized that parameter choice is crucial and called for more subjective testing.
  • ANFIS was also used to identify relationships between QoS parameters affecting QoE and overall perceived QoE.

3.5. QoE model using Crowdsourcing for subjective tests

Crowdsourcing-based subjective testing supported a YouTube-specific QoE model linking user ratings to network-induced video stalling. Stalling dominated reported QoE effects, but anonymous crowdsourced responses introduced reliability concerns.

  • The study combined a YouTube QoE model based on user ratings and stalling events with a cost-efficient, flexible crowdsourcing methodology.
  • QoE was primarily influenced by video stalling, while internet usage level, age, and video content type showed no significant impact.
  • Users may tolerate one stalling event below three seconds per clip but tend to be highly dissatisfied with two or more events.
  • Anonymous crowdsourcing can reduce reliability because participants may submit invalid or low-quality work for payment.The study therefore calls for methods that test subject trustworthiness and response quality.
  • A resource-arbitration framework similarly treats QoE as a combined function of application- and network-layer QoS rather than isolated parameters.It dynamically adjusts delay, jitter, and packet loss using subjective feedback.

3.7. QoE model considering equipment and environment factors

The reviewed QoE model incorporates equipment and environmental conditions alongside network parameters, adapting signal quality to users’ circumstances. Results indicate improved satisfaction for resource-constrained users, while further optimisation and broader testing remain necessary.

  • QoE model inputs: The QoE video model combines environmental and equipment factors with throughput and bit rate.The user’s equipment tests hardware and environmental conditions before network signal quality is adjusted.
  • Observed effects: The QBS algorithm improved satisfaction, especially for resource-constrained users.The model adjusts network signal quality according to equipment and environmental test results.
  • Observed effects: Surrounding interference, including light, noise, and shaking, produced different QoE levels for users.The reported model links higher-quality signals to high-end devices in low-interference environments, and the reverse condition otherwise.
  • Scope and caveat: The QBS algorithm requires further optimisation and deeper QoE testing across scenarios with additional QoS parameters.
  • Related assessment: The combined RST and CCA approach found that different video content types require different levels of QoS support.It assessed the combined impact of application QoS, network QoS, and content characteristics on QoE.

4. ANALYTICAL REVIEW

The analytical review compares how selected QoE/QoS models measure quality, map QoS inputs to QoE outputs, and represent influencing factors. It finds widespread use of subjective scores and coefficient-based mappings, including artificial optimisation techniques.

  • Evaluation approaches: Most analysed models collect subjective user measurements through the TUQ approach, commonly using MOS scores.
  • Evaluation approaches: Other models use objective technical measurements, while some combine subjective and objective variables or collect qualitative user opinions.
  • Measurement interpretation: Objective measures can include user behaviour, such as task duration and mouse-click counts, rather than only technology-derived measurements.
  • Parameter mappings: QoS parameters serve as inputs to nonlinear mapping functions whose outputs are MOS or objective metrics such as PSNR, VQM, and SSIM.
  • Parameter mappings: Most reviewed models use coefficient methods, with artificial optimisation techniques learning coefficients to find an optimal QoE fit.
  • QoE factors: QoE influence factors are classified across human, system, context, application, resource, and user dimensions.

Factors influencing Video QoE

Video QoE depends on network impairments, content characteristics, and additional factors that are unevenly covered by reviewed studies. The findings include specific effects of stalling, temporal structure, network design, and adaptive resource selection.

  • Network factors: Network impairments such as packet loss, delay, and jitter were investigated by all reviewed models.Burst losses, packet prioritisation by type, and dejittering buffer size received comparatively little attention.
  • Content factors: Video content type was the second most studied QoE influence factor, with 14 studies covering it.Packet loss dominated for slow-moving clips, whereas fast-moving football also depended on video bit rate.
  • Coverage limitations: Privacy, audiovisual interaction, user interface, quality awareness, cost, and last-mile equipment or environment were rarely investigated.
  • Main findings: A proposed pre-encoding scheme predicts encoded-service video quality and applies to MPEG-based sequences with a specific GOP structure.
  • Main findings: The reviewed findings indicate that network design protecting video flows may influence perceived quality more strongly than commonly controllable MPEG-4 codec parameters.
  • Main findings: QoE models were reported as suitable for real-time IPTV monitoring and resource adaptation, with over 10,000 flows handled in < 4ms.
  • Main findings: Rebuffering frequency and temporal structure were identified as important QoE determinants, rather than spatial artifacts alone.
  • Main findings: One stalling event per clip was tolerated when its duration stayed below 3 s, while internet usage, age, and content type showed no significant impact.

5. DISCUSSION

Existing QoS/QoE models pursue prediction in either direction, but interactions among QoS parameters and QoE remain insufficiently quantified. The reviewed literature indicates that intelligent optimization can improve mapping accuracy, while parameter selection, data volume, and contextual factors remain important.

  • Existing models either predict QoE from QoS parameters or identify suitable QoS for a desired QoE.
  • No standardized methodology directly and quantitatively maps QoS to QoE, leaving many existing models as partial prediction approaches.
  • Mapping approaches include top-down QoE-driven models, bottom-up QoS-driven models, and combinations of both.
  • Intelligent optimization models outperform other reviewed methods by learning nonlinear QoS–QoE relationships and fitting QoE metrics to QoS parameters.The paper notes that larger video databases and careful parameter choice are needed for greater prediction accuracy.
  • QoE/QoS correlation requires identifying interacting QoE factors, conducting quantitative subjective measurements, and deriving relevant QoS parameters and quality thresholds.Contextual information may support contextualized QoE but introduces security and privacy issues.

6. CONCLUSION

The paper reviews QoE/QoS correlation models as a challenging route toward optimizing multimedia services. It finds that combined quantitative subjective and objective measurements, together with optimized coefficients and intelligent techniques, better capture nonlinear QoS–QoE relationships, although broad application coverage remains unresolved.

  • The review analyzes QoE/QoS models by comparing measurement approaches, quality metrics, QoS parameters, and factors affecting QoE.
  • Models combining quantitative subjective and objective variables better reflect QoE complexity than models using qualitative subjective data or technical measurements alone.Quantitative data enables statistical descriptions and combined analysis of subjective and objective variables.
  • Optimized coefficients provide a better fit than constant coefficients for mapping nonlinear QoS parameters to QoE.Reviewed artificial techniques include ANN, Fuzzy Logic, DT, and SVM methods.
  • Most reviewed models remain partial, application-specific approaches with different computational and operational requirements.A model capturing how quantitative QoE is directly affected by QoS parameters is still missing.
Loading 1306.0221v1…