Title: Robust Hybrid Beamforming via GNNs: A Framework for Score-Based Channel State Information Generation and Denoising

Abstract:

High-quality Channel State Information (CSI) is a fundamental requirement for effective Hybrid Beamforming (HBF). Nevertheless, acquiring high-resolution CSI poses significant difficulties in real-world wireless communication environments. To overcome these limitations, this study introduces a robust HBF approach leveraging Graph Neural Networks (GNNs) and score-based generative models to operate effectively under imperfect CSI conditions.

Our methodology begins with the development of the Hybrid Message Graph Attention Network (HMGAT), which refines both node and edge features via distinct node-level and edge-level message passing mechanisms. Subsequently, we construct a Bidirectional Encoder Representations from Transformers (BERT)-based Noise Conditional Score Network (NCSN). This component is designed to model the distribution of high-resolution CSI, thereby enabling data augmentation and CSI generation that enhance the performance of the HMGAT. Finally, we introduce a Denoising Score Network (DSN) framework, specifically instantiated as DeBERT. This system is capable of removing noise from imperfect CSI across varying degrees of channel error, thus supporting robust HBF operations.

Evaluations conducted on the DeepMIMO urban dataset confirm that our proposed models exhibit superior generalization, scalability, and resilience across diverse HBF scenarios, whether operating with perfect or imperfect CSI.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC