Sensing-Prior Dual-Domain Graph Attention for User Scheduling in Wi-Fi 7 MU-OFDMA

Abstract

Multi-user OFDMA in IEEE 802.11be (Wi-Fi 7) suffers from residual inter-user interference caused by residual carrier frequency offset, transmitter spectral leakage, and antenna-domain spatial coupling, even after the access point (AP) enforces standard-compliant frequency synchronization. This residual interference becomes the dominant SINR ceiling under dense uplink admission and high-order modulation. Conventional MMSE-based receivers depend on accurate channel state information and degrade under residual impairments, while existing learning-based schedulers treat the per-user interference structure as a black box and represent it through complete or random graphs whose edge weights carry no physical meaning. To address these limitations, we propose to repurpose the by-products of IEEE 802.11bf Wi-Fi sensing— angle-of-arrival, delay spread, and residual CFO—into a sensing-prior dual-domain interference graph for AP-side scheduling, in which each directed user-pair edge weight is the product of a closed-form array-steering spatial term and a leakage-intercarrier-interference frequency term integrated over each resource unit. A physics-aware graph attention network (PA-GAT) injects the log edge weight into its attention logits and outputs joint per-user log signal/log interference power, from which a predicted SINR drives top-K' user scheduling. Simulation results under TGax indoor channels, cross-scenario testing, and corrected-CFO evaluation show that joint S/I based scheduling approaches the per-user Oracle-SINR while outperforming Random and I-only admission rules.