QoE Modeling: From QoS to Perception

• Quality of Experience (QoE) is defined as the study of user-perceived service quality, linking measurable QoS metrics like delay, jitter, and packet loss with subjective factors such as delight or annoyance.
• QoE modeling employs both retrospective and continuous methods, using metrics like MOS and time-varying quality traces to capture dynamic user satisfaction in multimedia and streaming services.
• Methodologies range from simple regression and logistic mappings to advanced state-space, deep learning, and control-theoretic models, enabling precise prediction and operational optimization of user experience.

Quality of Experience (QoE) modeling is the formal study of how users perceive the quality of a service and how that perception can be inferred from measurable variables. In multimedia, telecom, and streaming systems, QoE is usually framed as a user-centric construct that depends on perceived quality, satisfaction, expectation, and context, while Quality of Service (QoS) denotes objective network or system behavior such as delay, jitter, packet loss, throughput, bitrate, or buffer dynamics. Depending on the application, QoE may be represented by Mean Opinion Score (MOS), a continuous time-varying quality trace, startup delay and interruption probability, or broader constructs such as the degree of delight or annoyance and its behavioral consequences (Alreshoodi et al., 2013, Mitra et al., 2014, Koniuch et al., 2023).

1. Conceptual foundations

QoE modeling is built on a persistent distinction between technical performance and human judgment. In multimedia services, QoS refers to deterministic network or application behavior—minimum packet loss, acceptable delay, limited jitter, sufficient bandwidth, codec configuration, frame rate, resolution, and related variables—whereas QoE is the subjective, human-centric outcome that ties together perception, expectations, and experience of the application and network performance. The survey literature makes the generic relationship explicit as $QoE = f(QoS_1, QoS_2, \dots, QoS_n)$, but also emphasizes that this mapping is typically nonlinear and mediated by content, device, environment, and user factors (Alreshoodi et al., 2013).

Later reviews broaden this view by treating QoE as dependent not only on underlying QoS but also on a person’s preferences, expectations, experience, behavior, cognitive abilities, the service’s attributes, and the surrounding environment. In telecom and OTT scenarios, QoE is explicitly defined as the “degree of delight or annoyance of the user,” shaped by system-level factors, contextual factors, and human factors or perceptual dimensions. This shifts QoE modeling away from pure network inference toward multidimensional user-state estimation (Mitra et al., 2014, Nayar et al., 30 Apr 2025).

A stronger causal interpretation appears in structural treatments of video QoE. There, QoS and content jointly affect the Quality of Multimedia Signal (QoMS), QoMS influences Perceived QoMS (PQoMS), content and PQoMS jointly shape the Degree of Delight or Annoyance (DoA), and DoA drives behavior. This formulation is significant because it separates objective signal quality from subjective perception and further separates perceived quality from the downstream behavioral outcome that many services ultimately care about (Koniuch et al., 2023).

2. Measurement targets and QoS–QoE mappings

MOS remains the dominant scalar target in QoE modeling, but its role varies. In classical multimedia studies it is often a 5-point subjective scale, while some streaming datasets use continuous scores on $[0,100]$ or time-varying traces. Subjective methodologies include MOS panels, DSCQS, continuous slider-based annotation, and crowdsourcing; objective proxies include VQM, PSNR, SSIM, STRRED, NIQE, PESQ, and P.1203-derived scores. The literature repeatedly stresses that objective metrics are not themselves QoE, but inputs to a mapping whose validity depends on the task and context (Alreshoodi et al., 2013, Mitra et al., 2014).

Parametric mappings range from simple linear regression to explicitly nonlinear forms. Representative examples include linear mappings such as $MOS_p = a_1 + a_2 \cdot VQ$, logistic, cubic, logarithmic, power-law, and exponential-sum mappings, as well as the IQX hypothesis

$QoE = a \cdot \exp(-\beta \cdot QoS) + \gamma,$

which models exponential decay of QoE as QoS disturbance increases (Alreshoodi et al., 2013). In VoIP-related formulations, objective QoS values can be transformed to MOS through E-model style impairment equations and an $R$-to-MOS conversion, providing an explicit route from delay, codec and loss, and jitter to a subjective scale (Daengsi et al., 2023).

Decision-theoretic and utility-based mappings represent a second major tradition. For 5G real-time and interactive services, one model defines

$$MOS_{\text{Overall}} = 0.55 \cdot MOS_{\text{Loss}} + 0.25 \cdot MOS_{\text{Delay}} + 0.20 \cdot MOS_{\text{Jitter}},$$

with weights derived by Analytic Hierarchy Process (AHP) from pairwise comparisons by seven evaluators, indicating that packet loss is perceived as more harmful than delay or jitter in that usage context (Daengsi et al., 2023). In media streaming over heterogeneous access networks, QoE is modeled not by MOS but by the pair $(D,\epsilon)$, where $D$ is initial buffering delay and $\epsilon$ is tolerated interruption risk; for a single-server model the interruption probability obeys

$p(D) = e^{-I(R)D},$

which directly links arrival rate and startup buffering to playback reliability (ParandehGheibi et al., 2012). For mobile YouTube, a utility function maps initial buffering time, rebuffering frequency, and mean rebuffering time to MOS, and empirical calibration shows that wireless users can be more indulgent than the original wired-scenario model predicts (Gomez et al., 2014).

3. Streaming-specific and temporal QoE models

Streaming applications introduced a decisive change in QoE modeling: user experience is not only session-level but also dynamic. Retrospective models estimate a single end-of-session score from impairments accumulated over time, while continuous models predict QoE as a time series. In HTTP adaptive streaming, this requires accounting for bitrate fluctuations, rebuffering events, their timing, and memory effects. Video ATLAS is a representative retrospective model: it combines an objective video-quality feature $[0,100]$0, rebuffering features $[0,100]$1 and $[0,100]$2, a memory feature $[0,100]$3, and an impairment-duration feature $[0,100]$4, then learns a regression from these features to subjective QoE. Its central contribution is to show that distortions, rebuffering, and memory must be modeled jointly rather than through isolated quality or stall statistics (Bampis et al., 2017).

Continuous QoE models make the temporal dependency explicit. LSTM-QoE treats streaming QoE as a nonlinear, non-Markovian stochastic process driven by Short Time Subjective Quality (STSQ), Playback Indicator (PI), and time elapsed since the last rebuffering event $[0,100]$5. Using a two-layer, 22-unit-per-layer LSTM and a dense output layer, it predicts per-second QoE traces and outperforms NARX, SVR-QoE, Hammerstein–Wiener, and NLSS-QoE on several continuous QoE datasets, showing that long- and short-term memory are essential for modeling hysteresis and recovery (Eswara et al., 2018).

Convolutional sequence models offer a computational alternative to recurrent architectures. CNN-QoE uses an initial causal convolution to capture local temporal structure and a stack of dilated causal convolutions to model longer-range dependencies over an 8-second receptive field, with STSQ, PI, number of rebuffering events, and time since last impairment as inputs. On LFOVIA, LIVE Mobile Stall II, and LIVE Netflix, it matches or exceeds state-of-the-art continuous predictors while achieving lower inference latency on both PCs and mobile devices, which is consequential for edge deployment and client-side control (Duc et al., 2020).

Live streaming introduces additional distortions that do not fit standard VoD assumptions. TaoLive QoE provides the first live video streaming QoE dataset, with 42 source videos and 1,155 distorted ones including compression, stalling, frame skipping, variable frame rate, and accelerated playback after stalls. Tao-QoE addresses this setting with a retrospective end-to-end model that reconstructs stalls via PTS manipulation, extracts multi-scale semantic features with Swin Transformer, motion features with PWC-Net and a 3D ResNet, and fuses temporal differences to predict a scalar QoE score without relying on QoS features. The reported benchmarks show that many VoD-oriented QoE models struggle on live content, particularly when frame skipping and variable frame rate are present (Zhu et al., 2024).

4. State-space, queueing, and control-theoretic formulations

A parallel research line treats QoE as a dynamical system or stochastic control problem. The nonlinear state-space model NLSS-QoE defines a static nonlinear mapping from STSQ, PI, and $[0,100]$6 to transformed inputs $[0,100]$7, then evolves a discrete-time state vector according to

$[0,100]$8

With model order $[0,100]$9, the resulting state-space is compact, and the model is shown to be completely state controllable and observable. This is important because it turns QoE prediction into a system-theoretic object that can, in principle, be embedded in feedback control or model-predictive adaptation (Eswara et al., 2018).

Analytical queueing models push this idea further by defining QoE directly in terms of starvation, startup delay, and bitrate adaptation. In technology- and cost-heterogeneous streaming, QoE is the trade-off between initial waiting time and interruption probability, while expected costly-server usage becomes the optimization objective. The system is formulated as a continuous-time MDP with a probabilistic constraint, and the Hamilton–Jacobi–Bellman equation yields explicit threshold policies in the fluid approximation. This makes QoE modeling operational in the sense of optimal control rather than regression alone (ParandehGheibi et al., 2012).

For long-range-dependent traffic, a Markov Modulated Fluid Model approximates bursty arrivals and yields closed-form expressions for startup delay and starvation distributions via PDEs and Laplace transforms. QoE is summarized by an additive cost

$MOS_p = a_1 + a_2 \cdot VQ$0

combining number of starvations, startup delay, and bitrate degradation with user preference weights. This makes explicit that QoE optimization is often a trade-off among multiple objective impairments rather than a single score (Ye et al., 2014).

Large-scale measurement of user behavior adds another layer. In wireless video streaming, measured viewing times fit a two-phase hyper-exponential distribution, implying short and long views with separated time scales. This supports a two-class CTMC and QoE analysis under discriminatory processor sharing, where starvation metrics for short and long views respond differently to scheduling weights. The result is a flow-level QoE model tied directly to network scheduling policy, plus an online Bayesian approach to infer viewing-time class from minimal provider-side information (Xu et al., 2016).

5. Generative, personalized, operator-side, and cross-domain extensions

Recent work extends QoE modeling beyond deterministic regressors. A lightweight generative framework first vector-quantizes multivariate QoS and VQA features into discrete symbols and then applies a first-order HMM to model latent QoE states. This VQ–HMM pipeline is explicitly positioned as a model of $MOS_p = a_1 + a_2 \cdot VQ$1, supports probabilistic inference on unseen sequences, and delivers accuracy $MOS_p = a_1 + a_2 \cdot VQ$2 with inference latency $MOS_p = a_1 + a_2 \cdot VQ$3 s, compared with $MOS_p = a_1 + a_2 \cdot VQ$4 for HMMs on naïvely binned data and roughly $MOS_p = a_1 + a_2 \cdot VQ$5 s latency for LSTM in the reported setup. The contribution is not only efficiency but also interpretability: transition and emission matrices expose temporal QoE structure (Nayar et al., 30 Apr 2025).

Operator-level QoE aggregation introduces a dual-layer design. One layer predicts objective MOS from network KPIs using a Random Forest trained to emulate the ITU-T P.1203 reference implementation, reaching $MOS_p = a_1 + a_2 \cdot VQ$6 and “97% parity” with the reference model. The other layer derives subjective MOS from live-stream comments through semantic filtering and deterministic LLM scoring. The resulting dataset contains 47,894 comments, about 34,000 of which are identified as QoE-relevant, and supports ISP-level aggregation via

$MOS_p = a_1 + a_2 \cdot VQ$7

This is a notable shift from user-session prediction to operator-side monitoring and anomaly detection (Panahi et al., 1 Jun 2025).

Personalization enters through demographic augmentation. A 5G video-streaming framework starts from 450 sessions, defines six synthetic user profiles with different sensitivities to rebuffering, quality, bitrate variability, and consistency, and expands the dataset to 2,700 samples by profile-specific MOS transformation plus Gaussian noise. Classical and deep tabular models are then trained on continuous MOS in $MOS_p = a_1 + a_2 \cdot VQ$8. On the augmented data, Random Forest attains RMSE $MOS_p = a_1 + a_2 \cdot VQ$9, MAE $QoE = a \cdot \exp(-\beta \cdot QoS) + \gamma,$0, and $QoE = a \cdot \exp(-\beta \cdot QoS) + \gamma,$1, while TabNet reaches strong correlation and is emphasized for feature selection and attention-based interpretability (Ahmed et al., 14 Dec 2025).

QoE modeling has also been generalized beyond multimedia. QoNext adapts QoE principles to foundation-model interaction, treating user experience as a function of content quality and service quality. Its five core factors are information density, content accuracy, output speed, latency position, and latency duration; subjective outcomes include Overall Impression, Content Quality, and Perceived Responsiveness. Predictive models achieve SRCC values around $QoE = a \cdot \exp(-\beta \cdot QoS) + \gamma,$2 for overall QoE, and ablation shows that content accuracy is the dominant factor while speed is the main secondary one. This suggests that the QoE paradigm can be transferred from streaming systems to interactive AI services without abandoning its core structure of subjective data collection, parameter control, and regression to MOS-like outcomes (Guo et al., 26 Sep 2025).

6. Methodological debates, validation practice, and open problems

A recurrent methodological divide is between bottom-up and top-down modeling. Bottom-up approaches enumerate many Influential Factors and Perceptual Dimensions, then apply factor analysis, regression, or machine learning to discover associations. Top-down approaches argue that this often obscures causal structure, ecological validity, and practical interpretability. The proposed structural model for video QoE therefore begins from a minimal causal skeleton—QoS and Content to QoMS, QoMS to PQoMS, Content and PQoMS to delight or annoyance, and delight or annoyance to behavior—and treats additional variables as controlled extensions rather than unconstrained correlates (Koniuch et al., 2023).

Validation practice also remains heterogeneous. Some studies rely on tightly controlled lab protocols such as ACR and continuous slider annotation; others use large-scale logs, field measurements, or crowdsourcing; newer systems incorporate semantic comment analysis or synthetic demographic profiles. This diversity is productive, but it complicates cross-study comparability. Reviews repeatedly note that many existing models are partial, service-specific, and weak in their treatment of device, environment, user heterogeneity, and context, while also warning that MOS is an ordinal scale that is often manipulated as if it were interval without sufficient justification (Alreshoodi et al., 2013, Mitra et al., 2014).

Several open problems recur across the literature. Small-sample subjective calibration remains common, as in the AHP-based 5G model built from seven evaluators, which the authors explicitly describe as an early conceptual-validation stage (Daengsi et al., 2023). Static weights may not transfer across service types, devices, or user groups; synthetic augmentation improves robustness but does not replace real-user demographic data (Ahmed et al., 14 Dec 2025). Live-streaming models remain challenged by variable frame rate and frame skipping, while operator-side systems must reconcile objective telemetry with platform-derived subjective signals (Zhu et al., 2024, Panahi et al., 1 Jun 2025). More generally, the literature points toward hybrid models that combine probabilistic temporal structure, perceptual feature extraction, richer contextual sensing, and causal interpretation. A plausible implication is that future QoE modeling will be less about a single universal MOS predictor than about interoperable models specialized by service type, temporal granularity, and deployment layer.