ResLearn-XR is a residual learning framework designed for predicting extended reality (XR) network traffic and estimating Quality-of-Experience (QoE) risk. It employs a two-stage temporal learning structure, combining a base sequence prediction model with task-specific residual components.
The residual stages operate in different spaces: the value space for continuous XR traffic forecasting and the logit space for probabilistic QoE risk estimation. A Data Descriptor Algorithm (DDA) converts packet-level observables into frame-timing-aware descriptors, aiding encrypted traffic analysis.
An XR Traffic-QoE dataset pairs traffic traces with user-reported QoE labels, supporting model training and evaluation. ResLearn-XR reduces SMAPE by up to 17.84% for traffic prediction and up to 87.8% for QoE risk estimation compared to single-stage baselines, which can benefit engineers managing XR network models.
Source: https://arxiv.org/abs/2609.04493