Adaptive GNN-based Proportional-Fair Scheduling in MIMO Networks for Non-stationary Channels

We study adaptive machine learning (ML)-based Proportional Fair (PF) scheduling for MIMO networks under non-stationary channel conditions, which are increasingly prevalent in dense MIMO deployments with complex scattering environments. Traditional PF schedulers face the trade-off limitations between utility and latency, whereas conventional ML-based schedulers, although balancing utility and latency better, degrade under changing channel distributions. We propose a Node-and-Edge Attention Graph Neural Network (NEA-GNN) that exploits both MIMO network structure and PF scheduling criteria. NEA-GNN employs attention-based message passing to jointly capture inter-user interference patterns and per-user metrics, enabling high transmission fairness and throughput while achieving low adaptation overhead, improving long-term utility under dynamic channels. Experiments over simulated and real-world channel measurements demonstrate that NEA-GNN is more sample-efficient than conventional ML models and consistently outperforms existing PF schedulers in non-stationary MIMO environments.