Bayesian Sample Augmentation in Graph Neural Networks for Radio Map Recovery
Weihua Luo, Fasheng Zhou · 2024
Through the radio map presented by the signal strength at observed points, the signal distributions such as signal attenuation of an area can be obtained. However, it is practically difficult to collect adequate samples from dense urban environment. Therefore, we aim to address this research issue of recovering radio map via randomly sparse observations. Particularly, different from existing works, we adopt a Bayesian optimization (BO) method to screen the training samples based on the graph attention network (GAT). Numerical simulations shows that the proposed BOGAT can achieve more accurate performance with even sparser observations as compared to common existing GAT models.