Data-Enhanced GNN via Simplex Vectors in Radio Map Prediction with Sparse Samples
Yingjin Pan, Fasheng Zhou · 2025
Radio map constituted by the received-signal strength (RSS) at observed points can be leveraged to enhance the services of either communications or target sensing. In this paper, we aim to predict the overall radio map via only partial random RSS samples of the observed points. To facilitate the prediction, we adopt a graph neural network (GNN) framework on which a learning structure is devised. Specifically, to enhance the efficiency of the learning approach, we devise a data-enhanced method based on simplex theory where the convex simplex vectors are first searched and found, based on which an extended convex training data-set for the prediction learning is deliberately constructed and it is convex. Then, this convex data-set is trained by a purposely designed GNN to fulfill the aim of prediction. Numerical studies are performed to verify the efficiency of the proposed GNN learning structure.