SAMS-GNN: Self-Adaptive Multi-Scale Graph Neural Network for Multi-Band Spectrum Prediction

Abstract

The rapid advancement in wireless communication technology has led to a high demand for spectrum resources, causing a scarcity of available spectrum. However, current spectrum prediction methods primarily focus on temporal correlations, neglecting the influence between different frequency bands and failing to address the need for multi-time-scale predictions in real-world scenarios. In this paper, we propose a self-adaptive multi-scale graph neural network (SAMS-GNN) for enhanced multi-band spectrum prediction. SAMS-GNN incorporates multi-time-scale data inputs into the network to capture dependencies across various temporal scales, simultaneously preserving the global and local dependencies of the spectrum data. A self-adaptive graph structure is utilized to dynamically learn the shifting dependencies between frequency bands, while the Gated Recurrent Unit (GRU) is employed to distill temporal dependencies. Different time scales share the same network parameters. Furthermore, we introduce a scale fusion module to weigh the contributions of different time scales, enhancing the collaboration between scales. Experiments on two real-world spectrum datasets demonstrate that SAMS-GNN outperforms existing spectrum prediction methods in prediction performance.