QoE Prediction

The core of QoE prediction is a learned mapping from objective network KPIs to subjective user experience scores.

• Input Features: Packet loss rate, jitter, latency, throughput, and Channel Quality Indicator (CQI) values
• Output Metric: Mean Opinion Score (MOS) on a 1-5 scale, or application-specific metrics like video stall duration and time-to-first-byte
• Non-Linear Relationship: A small increase in packet loss can cause a catastrophic drop in QoE due to video buffer starvation, making simple threshold-based approaches inadequate
• Modern systems use deep neural networks to capture these complex, non-linear dependencies between network conditions and perceptual quality

Different applications have fundamentally different tolerance profiles for network impairment. A single QoE model cannot serve all traffic types.

• Video Streaming: Highly sensitive to throughput fluctuations and stall events; adaptive bitrate (ABR) algorithms can mask minor issues but create step-function degradation
• Web Browsing: QoE correlates with Page Load Time and Speed Index, which depend on both throughput and round-trip time (RTT)
• Cloud Gaming: Requires ultra-low latency (<30ms) and zero packet loss; jitter is often more critical than average throughput
• VoLTE/VoNR: Predicts listening quality using POLQA or PESQ algorithms, sensitive to packet loss concealment effectiveness
• Effective systems use traffic classification as a pre-processing step to route telemetry to the correct application-specific prediction model

User QoE is not determined by a single instantaneous network event but by the sequence and duration of impairments over time.

• Recency Effect: A stall at the end of a video session has a disproportionately negative impact on remembered quality compared to one at the beginning
• Stall Frequency vs. Duration: Multiple short stalls can be more annoying than a single longer stall, requiring models to capture event patterns
• LSTM and Transformer architectures are used to model these long-range temporal dependencies in the KPI time-series
• Attention mechanisms allow the model to weight the importance of past network events when predicting current perceived quality
• This temporal context is what separates true QoE prediction from simple QoS threshold alerting

Training a QoE prediction model requires labeled data that captures genuine human perception, which is expensive and difficult to obtain at scale.

• ITU-T P.800/P.808: Standardized methodologies for conducting subjective quality assessments with human raters in controlled environments
• Crowdsourced Labeling: Large-scale QoE data collection using platforms like Amazon Mechanical Turk, though with higher variance
• Objective Proxy Metrics: SSIM, VMAF, and POLQA are algorithmic estimators of perceived quality used to generate pseudo-labels when human annotation is infeasible
• Transfer Learning: Models pre-trained on large-scale objective datasets are fine-tuned with a smaller corpus of subjective human ratings
• The ground truth gap—the mismatch between predicted and actual human perception—remains the central challenge in QoE prediction research

The ultimate value of QoE prediction is closing the loop—using forecasts to trigger preemptive network actions that prevent user experience degradation.

• Predictive Bitrate Steering: If a video session is predicted to stall, the network can proactively lower the adaptive bitrate before the buffer empties, avoiding a visible stall event
• QoE-Aware Load Balancing: Handover decisions are optimized not just for load equalization but for maximizing predicted aggregate MOS across all users
• Slice Resource Reallocation: In 5G network slicing, resources are dynamically shifted to a slice experiencing predicted QoE degradation to maintain its SLA
• Near-RT RIC Integration: QoE prediction models run as xApps on the Near-Real-Time RAN Intelligent Controller, executing control loops on a 10ms to 1s timescale
• This transforms QoE prediction from a passive monitoring tool into an active QoE assurance system

Advanced QoE prediction systems fuse heterogeneous data sources beyond raw network KPIs to improve forecast accuracy.

• Device-Side Metrics: Buffer level, playback state, and CPU load from the User Equipment (UE) provide direct insight into the playback pipeline
• Application-Layer Signals: HTTP Adaptive Streaming (HAS) manifest requests and chunk download times reveal the ABR algorithm's decision state
• Contextual Features: Time of day, location, device type, and subscription tier add predictive power by capturing usage context
• Radio Layer Telemetry: CQI reports, MIMO rank, and HARQ retransmission counts from the PHY/MAC layers provide early warning of impending channel degradation
• Fusing these modalities requires careful time alignment and feature engineering to ensure all signals correspond to the same user session

The fundamental subjective metric that QoE prediction models aim to forecast. Originally a human-rated scale from 1 (bad) to 5 (excellent), modern systems estimate MOS algorithmically by mapping network KPIs to perceptual quality.

• Full-Reference models: Compare original signal to received signal (PSNR, SSIM)
• No-Reference models: Estimate quality from received signal alone using ML
• Parametric models: Predict MOS directly from IP-layer metrics like packet loss and jitter

A QoE prediction for video streaming might output: 'Predicted MOS: 3.8 — minor stalling expected.'

Unlike network-layer metrics (latency, throughput), these indicators directly measure the user's perceived service quality. QoE prediction models must map forecasted network conditions to these application-specific outcomes.

• Video: Stalling duration, stalling frequency, initial buffering time, playback resolution switches
• Web Browsing: Page Load Time (PLT), Time to Interactive (TTI), Speed Index
• Cloud Gaming: Frame rate drops, input lag, rendering quality degradation
• VoLTE/VoNR: Mouth-to-ear latency, jitter buffer underruns, Mean Opinion Score for Listening Quality (MOS-LQ)

Predicting a 2-second stall event 5 seconds before it occurs enables preemptive bitrate switching.

The mathematical or ML-based transfer functions that convert predicted network KPIs into estimated user experience scores. These are the core inference engines in a QoE prediction pipeline.

• Parametric Models: ITU-T G.107 (E-model) for voice, using delay and packet loss as inputs
• Machine Learning Regression: Random Forest or Gradient Boosting trained on labeled QoE datasets
• Deep Neural Networks: CNNs or LSTMs that learn non-linear relationships between multivariate KPI sequences and MOS
• Transfer Learning: Adapting a mapping function trained on one service type (e.g., video) to another (e.g., cloud gaming)

Example: A predicted throughput drop to 2 Mbps maps to a 70% probability of 240p video resolution, triggering a QoE alert.

The International Telecommunication Union provides the standardized frameworks that define how QoE is modeled, measured, and managed in operational networks.

• ITU-T P.1203: Standard for parametric bitstream-based QoE assessment of HTTP Adaptive Streaming (HAS), outputting per-segment quality scores
• ITU-T G.107 (E-model): Computational model for conversational voice quality using transmission impairment factors
• ITU-T P.800: Defines the Absolute Category Rating (ACR) methodology for collecting subjective MOS scores
• ITU-T G.1030: Framework for estimating web browsing QoE from network-level measurements

These standards ensure vendor-neutral, repeatable QoE prediction that can be audited and compared across deployments.

QoE prediction directly informs Adaptive Bitrate algorithms to preemptively switch video quality before a network bottleneck causes a buffer underrun and visible stalling.

• Buffer-Based ABR: Uses current buffer occupancy; enhanced by predicted future throughput
• Hybrid ABR: Combines buffer state with forecasted channel conditions from QoE prediction engine
• Model Predictive Control (MPC): Optimizes bitrate ladder selection over a receding horizon using predicted QoE as the cost function

A QoE-aware ABR controller receiving a 10-second throughput forecast can smoothly degrade from 1080p to 720p rather than experiencing a sudden stall event.

In 5G networks, QoE prediction enables slice-level resource orchestration that proactively guarantees experience SLAs rather than just network KPIs.

• Slice SLA Translation: Converts a '99% of users with MOS > 4.0' SLA into dynamic PRB allocation targets
• Cross-Slice Arbitration: When resources are scarce, QoE predictions prioritize slices where user experience is most at risk
• Predictive Slice Scaling: Forecasts QoE degradation across a slice and triggers preemptive capacity expansion

For a URLLC slice carrying haptic feedback, a predicted QoE violation triggers immediate resource preemption from an eMBB slice with non-critical buffered video.