gnn-based online beamforming

本文针对6G宽频系统中波束倾斜问题，提出使用图神经网络优化混合波束成形的方法，通过设计三种不同的GNN结构提高了通信性能和计算效率。

To address the low accuracy and high computational cost of radio signal propagation modeling in real-world scenarios, this paper proposes a data-driven coverage map generation method based on Graph Neural Networks (GNNs). The approach automatically constructs a heterogeneous graph from environmental imagery, jointly encoding spatial topology and ray-based propagation relationships, and learns propagation patterns end-to-end using sparse, real-world signal measurements as supervision. This work is the first to apply GNNs to empirical radio propagation modeling, eliminating reliance on traditional physics-based solvers. Experiments demonstrate that the method surpasses classical numerical solvers (e.g., FDTD) and heuristic models in coverage map reconstruction accuracy, achieves 10–100× faster inference, and exhibits strong generalization—enabling high-fidelity signal coverage prediction from only a few measurements. These advances significantly enhance the efficiency and practicality of wireless network deployment.

This work addresses the high computational complexity of conventional ray tracing methods, which hinders real-time wireless network modeling. The authors propose a novel approach that integrates the physical principles of ray tracing with graph neural networks by converting environmental point clouds into graph structures and leveraging neural message passing to efficiently infer propagation parameters such as signal strength in three-dimensional space. This method establishes the first learnable and generalizable digital twin model of radio environments, enabling joint training on both simulated and real-world measurement data. It achieves high prediction accuracy while significantly reducing inference time, making it well-suited for efficient modeling and prediction in complex 3D wireless scenarios.

To address beam squint—a frequency-selective distortion in analog beamforming that severely degrades performance in 6G terahertz (THz) wideband MIMO-OFDM systems—this paper proposes a graph neural network (GNN)-based hybrid beamforming method. We innovatively design a dual-type graph node model, wherein nodes explicitly represent analog and digital beamformer matrices, enabling low-overhead, robust, real-time adaptive optimization. The proposed scheme effectively suppresses beam squint across wide bandwidths, achieving spectral efficiency close to fully digital beamforming while reducing computational and memory overhead to only a small fraction of conventional approaches. Crucially, it delivers exceptional stability in spectral efficiency over wideband operation. This work establishes a scalable, low-complexity, and practically viable beamforming paradigm for THz communications.

本文利用图神经网络（GNN）从Sub-6 GHz信道状态信息中学习毫米波无蜂窝大规模MIMO系统的波束成形，以降低训练开销并提高下行总速率。

This work addresses the challenges in compact-aperture fluid antenna arrays, where channel-driven port placement often leads to port clustering, exacerbated mutual coupling, and uneven current loading. To overcome these issues, the authors propose an electromagnetically guided graph neural network framework that, for the first time, integrates electromagnetic constraints—such as mutual impedance and geometric layout—into graph-based learning. This approach jointly optimizes port configuration and current-domain beamforming strategies under a unified electromagnetic feasibility criterion. By doing so, it enables a controllable trade-off among communication rate, current distribution uniformity, and configuration latency, thereby significantly enhancing the overall performance of multiuser downlink transmissions.

This work addresses the performance degradation of graph neural networks (GNNs) when transferring across scales on wireless conflict graphs modeled as sparse random geometric graphs. It establishes, for the first time, theoretical performance bounds for GNN generalization across conflict graphs of varying sizes. By analyzing the structural approximation between random geometric graphs and deterministic grid graphs, the study reveals a theoretical connection between graph topological similarity and model generalization capability. Guided by this theoretical insight, the proposed method significantly outperforms existing baselines in link scheduling tasks, with empirical results on large-scale scenarios further validating the tightness and practical relevance of the derived theoretical bounds.

This work addresses the low energy efficiency experienced by cell-edge users in terrestrial cellular networks, which stems from path loss, shadowing, and inter-cell interference. To mitigate these challenges, the authors propose a high-altitude platform station (HAPS)-assisted cooperative beamforming architecture that leverages line-of-sight links between HAPS and ground base stations for data relaying. An innovative online optimization framework integrating graph neural networks (GNNs) is developed to effectively model the dynamic network topology and solve the non-convex energy efficiency maximization problem in real time. Experimental results demonstrate that the proposed approach significantly improves the 5th-percentile network energy efficiency, thereby substantially enhancing quality of service for edge users.

This work addresses the challenge of decentralized multi-channel transmit power allocation in resource-constrained mobile ad hoc networks (MANETs), where existing approaches struggle to achieve efficient and real-time optimization. The paper proposes MANET-GNN, the first framework to leverage graph neural networks (GNNs) for this setting, modeling network topology through message passing and enabling fast, unsupervised, decentralized power control. Notably, MANET-GNN operates without global channel state information, exhibits robustness to channel noise, and generalizes effectively across diverse network topologies and channel conditions. Experimental results demonstrate that the proposed method significantly improves throughput across various MANET scenarios and scales efficiently to large numbers of nodes and frequency bands, meeting the stringent real-time requirements of practical deployments.

This study systematically investigates the effectiveness boundaries and practical value of Graph Neural Networks (GNNs) across twelve application domains. Building upon a unified design space, it derives both spectral and spatial formulations of GNNs from first principles, analyzes their expressive power through the lens of the Weisfeiler–Leman test, and evaluates domain-specific graph construction strategies and architectural choices. The work establishes the first cross-domain analytical framework that disentangles genuine performance gains from baseline biases, uncovering common challenges such as heterophily, scaling effects, and deployment gaps. It clarifies the applicability limits of GNNs, highlights the discrepancy between leaderboard-topping models and deployable ones, and offers constraint-aware practical guidelines to address issues including oversmoothing, over-squashing, and distributional shifts.