Graph Frequency-Domain Attention Learning for Sensing-Aware Predictive Beamforming in RIS-Assisted ISAC Vehicular Networks

Reconfigurable intelligent surface (RIS)-assisted vehicle-to-infrastructure (V2I) integrated sensing and communication (ISAC) has emerged as a promising solution to enhance connectivity. Nevertheless, existing RIS-assisted ISAC designs for vehicular scenarios typically rely on instantaneous channel state information (CSI), which limits their effectiveness in highly dynamic traffic environments. In this paper, we consider a highway V2I scenario where the direct base station-vehicle link is blocked and a semi-passive RIS with embedded sensing elements is deployed to simultaneously sense multiple vehicles and assist downlink communication. In particular, ISAC signals are exploited to maximize the user communication rate under sensing-accuracy constraints. To address the resulting dynamic and interdependent decision-making problem, two learning-based predictive beamforming models are proposed. First, a three-stage learning-based framework is developed, integrating echo-based multi-target localization, graph-frequency-domain (GFD) trajectory prediction, and Cramér-Rao lower bound (CRLB)-aware beamforming and RIS configuration. Target positions are first obtained from echo signals, after which GFD attention mechanisms are employed to capture the spatiotemporal features of vehicles, and an unsupervised model is used to perform beamforming based on the predicted channels. Furthermore, to reduce design complexity and eliminate the need for intermediate supervision, we propose an end-to-end GFD attention network that directly maps historical multi-target echoes to the future-slot transmit beamforming vector and RIS phase-shift matrix using a unified CRLB-aware loss function. Simulation results based on a real vehicle trajectory dataset demonstrate that the proposed schemes consistently satisfy sensing-accuracy requirements while achieving downlink rates close to the perfect CSI upper bound. Moreover, they significantly outperform outdated-CSI and existing end-to-end baselines across a wide range of RIS sizes and transmit power budgets.