State-Aware Analysis, Forecasting, and Predictive Scheduling of Virtual Reality Network Traffic for Enhanced Quality of Experience

Abstract

The escalating demand for interactive Virtual Reality (VR) applications places unprecedented demands on underlying network infrastructures, requiring stringent Quality of Service (QoS) and Quality of Experience (QoE) to mitigate motion sickness and ensure immersive user experiences. A critical challenge lies in adapting network resource management to the highly dynamic and user-state-dependent traffic patterns unique to VR, where traditional prediction methods often fall short in accounting for variations induced by user mobility (e.g., stationary versus moving). Moreover, effectively translating window-level traffic forecasts into actionable, fine-grained scheduling decisions at the millisecond scale presents a temporal disconnect limiting the practical utility of predictive intelligence. This thesis addresses these challenges comprehensively. First, it introduces a meticulously captured and explicitly annotated VR network traffic dataset, derived from controlled, reproducible VR sessions. This dataset distinguishes itself by comprehensively recording both server-to-client and client-to-server communication flows from a custom-built Unity VR application, crucially labeling traffic segments according to distinct user mobility states: stationary and moving. The dataset design and methodology enable reproducible research in immersive networking; the dataset and reproduction code are publicly available at \url{https://github.com/bozliu/vr-traffic-prediction-sps/tree/36a2d4d73c80c658d3d9174dcc1e9a08e9fc3922/data/raw}. Leveraging this rich data, the research demonstrates a novel approach to VR user-state classification using a Random Forest classifier. We propose a sliding-window-based multi-scale (0.5/1.0/2.0 s) hybrid time-frequency feature framework, robustly centering and performing FFT on packet size sequences to extract Nyquist-normalized frequencies and L1-normalized amplitudes of top-N peaks. This is augmented with spectral centroid/flatness descriptors and an uplink/downlink direction bit to capture VR traffic's periodicity and short-term dynamics without Deep Packet Inspection (DPI). This approach significantly enhances classification accuracy, increasing it from a time-only baseline of \(\approx 69\%\) to \(\approx 94\%\) on the test set (93.74\% overall; Moving F1=0.915). Feature-importance analysis highlights direction context and frequency position/shape (Nyquist-normalized frequencies, spectral centroid/flatness) as most discriminative, with amplitudes providing complementary information.

Furthermore, the thesis presents a comprehensive evaluation of models—including a Random Forest regressor, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer architectures—for short-term forecasting of average packet size. This vital metric for proactive resource allocation consistently shows robust predictability at the 200 ms horizon (R\(^2 \approx\) 0.85--0.86; RF R\(^2\)=0.8586, GRU 0.8576, LSTM 0.8531, Transformer 0.8529), while very short horizons (\(\le\)80 ms) remain challenging. Finally, integrating these predictive capabilities, the thesis develops and validates an end-to-end prediction-aided hybrid Semi-Persistent Scheduling (SPS) and Proportional Fair (PF) framework. This framework seamlessly integrates 200ms VR traffic predictions into a 1ms MAC layer scheduler. Through extensive multi-user simulations, this approach is shown to significantly improve delay fairness for VR users by approximately 3.0-3.5\% and reduce tail latencies, all while maintaining full network throughput. The framework's robustness is verified across various stress scenarios, demonstrating particular efficacy under bursty and highly multiplexed traffic conditions. Crucially, its SPS user selection achieves a high Jaccard overlap of approximately 0.91 with an ideal oracle scheduler, providing strong evidence of the practical utility of predictive intelligence in strategic resource management. This work lays crucial groundwork for enhancing the immersive and responsive nature of future VR experiences through intelligent, adaptive, and QoE-centric network management. Virtual Reality, VR network traffic, traffic classification; traffic forecasting; user mobility state; time-frequency features; Random Forest; LSTM; GRU; Transformer; Semi-Persistent Scheduling; Proportional Fair scheduling; Quality of Experience; 5G/6G networks.