Multiantenna Channel Map Prediction with Missing Location Information using Contrastive Learning and Graph Neural Networks

A radio map estimation method for multiantenna statistical channel state information (CSI) is proposed, which can interpolate the CSI measurements across space. A practical challenge is that for some measurements, location information may not be readily available, due to, e.g., poor GPS reception or privacy concerns. To circumvent this difficulty, contrastive learning is employed to learn a metric embedding space, in which both CSI and location measurements can be related. We also incorporate local patches of building maps around sensor locations, when locations are available. Based on the embeddings, the relative proximity of all measurements, with or without location tags, is captured in a graph. A graph neural network (GNN) is then employed to perform map interpolation. The numerical tests verify the effectiveness of the proposed graph construction and map estimation methods.