ResLearn-XR Uses Residual Learning to Model XR Traffic and QoE Risk

Summary

The paper introduces ResLearn-XR, a residual-learning framework for forecasting extended-reality (XR) network traffic and estimating the risk of poor quality of experience (QoE). Its two-stage temporal design combines a base sequence-prediction model with task-specific residual components intended to adapt to bursty, non-stationary XR traffic. The traffic branch performs continuous prediction in value space, while the QoE branch estimates risk probabilistically in logit space. For the QoE branch, the authors introduce a Data Descriptor Algorithm that causally converts packet-level application-layer observables into frame-timing-aware descriptors for analyzing encrypted traffic. They also build an XR Traffic-QoE dataset pairing continuous traffic traces with session-level user-reported QoE labels. Across frame-count, frame-size, and inter-arrival-time prediction, ResLearn-XR reduces SMAPE by up to 17.84% relative to single-stage baselines. Its QoE-risk branch reduces SMAPE by up to 87.8% against the same type of baseline. The abstract does not specify the dataset size, the underlying base model, or the full experimental setup.