Abstract
The challenge of mitigating interference in Space-Air-Ground Integrated Networks (SAGINs) is exacerbated by the inherent channel uncertainty, which arises due to dynamic weather conditions, heterogeneous user deployment, and different altitude of transmitters. To tackle this problem, Rate-Splitting Multiple Access (RSMA) has been seen as a promising solution due to its robustness. However, conventional beamforming designs for RSMA often suffer from two major limitations: high processing delays and overfitting to specific channel conditions. When the channel conditions change, the performance of these predictors degrades significantly, limiting their effectiveness in dynamic environments. To address these challenges, we propose a novel Fast-Adaptive Predictive Beamforming (FA-PB) framework for RSMA in SAGINs. Unlike traditional predictive beamforming approaches that rely on fixed predictive models, FA-PB integrates a transfer-learning-based online learning mechanism. This innovative approach allows the predictor to dynamically adapt to new channel conditions with minimal computational overhead. FA-PB achieves this by leveraging few-shot Channel State Information at the Transmitter (CSIT) samples, enabling real-time updates and adjustments to the predictor. Consequently, FA-PB ensures that the beamforming process can rapidly adapt to fluctuating channel conditions, maintaining high levels of performance even in highly dynamic SAGIN environments. Extensive simulation results validate the superiority of the FA-PB framework, demonstrating its enhanced adaptability and improved beamforming performance in SAGINs.