Recurrent and Dynamic Models for Predicting Streaming Video Quality of Experience

Area of Science:

• Computer Science
• Machine Learning
• Network Engineering

Background:

• Streaming video is a major bandwidth consumer, facing challenges from limited bandwidth and unreliable networks.
• Video impairments like compression artifacts and rebuffering degrade the user's Quality of Experience (QoE).

Purpose of the Study:

• To develop accurate, continuous-time subjective QoE prediction models for video streaming.
• To explore the use of recurrent dynamic neural networks and time-series forecasting for QoE prediction.

Main Methods:

• Utilized recurrent neural networks (RNNs) and non-linear autoregressive (NARX) models for time-series QoE forecasting.
• Integrated diverse inputs including video quality scores, rebuffering data, and memory effects on human behavior.
• Developed forecasting ensembles by aggregating multiple models to improve prediction accuracy and reduce variance.

Main Results:

• The proposed models demonstrated improved prediction performance for subjective QoE.
• Prediction accuracy approached human-level performance in evaluating video stream quality.
• Ensemble methods reduced forecasting variance compared to single models.

Conclusions:

• Recurrent dynamic neural networks offer a robust approach for predicting subjective QoE in video streaming.
• Accurate QoE prediction can enable perceptually optimized resource allocation for enhanced streaming services.
• The developed models and evaluation metrics advance the field of real-time video quality assessment.