Transport-Aware QoE Modeling for WebRTC: Predicting Playback Freezes with Machine Learning

Abstract WebRTC is the de facto standard for browser-based real-time media delivery, increasingly deployed over QUIC, a UDP-based transport protocol offering low-latency stream multiplexing and improved congestion control. While QUIC is designed to optimize performance for interactive applications, its impact on application-level Quality of Experience (QoE) remains underexplored, particularly from the perspective of client-side telemetry. In this work, we investigate how the underlying transport protocol impacts the observability and prediction of video playback freezes using machine learning. We construct two labeled datasets from real-world WebRTC sessions across diverse network environments, with QUIC selectively enabled and disabled. Using client-side statistics, we train Random Forest and Multi-Layer Perceptron models to predict freeze events. We analyze the feature landscape via both model-based importance and SHAP interpretability methods. Our results show that QUIC shifts the dominant predictors of QoE degradation from raw packet loss and retransmission metrics toward media-layer jitter, buffering, and recovery-related indicators. These findings highlight the need for transport-aware QoE modeling, and offer practical insights for developers building real-time QoE analytics directly within the browser environment.