CK-GAT: A Framework for Few-Shot Positioning Based on Channel State Information

Accurate positioning is a prerequisite for Harmonized Communication and Sensing. Most existing fingerprint positioning methods require a large amount of data for training, which is a labor-intensive and time-consuming task. Consequently, the issue of few samples in positioning tasks is inevitable. To accurately position with a limited number of samples, a new few-shot positioning framework is proposed, named CK-GAT. Specifically, the framework initially extracts features from the limited labeled data through a meticulously designed neural network to construct high-dimensional Channel Knowledge (CK), which characterizes the features of the current environment. Subsequently, the Channel State Information (CSI) that need to be predicted are processed through the same neural network to extract features, which are then embedded into the CK. These features are enhanced by leveraging the environmental characteristics learned by Graph Attention Networks (GAT) to achieve precise positioning. Experimental results demonstrate that under the condition of limited labeled data, the proposed CK-GAT framework outperforms the current state-of-the-art and their semi-supervised variants, achieving an improvement in positioning accuracy by 0.85 m in outdoor scenarios and 0.19 m in indoor scenarios.