GraphLoc: Enhancing Fingerprint-Based Localization With Graph Representation Learning

For 6G inherent intelligent capability, deep-learning-based wireless localization will become a promising technology for offering commercial location-based services (LBSs). Classical deep neural networks (DNNs) have been designed to learn feature representation for localization tasks. However, due to the uncertainty of radio measurements in complicated wireless propagation, the existing solution has achieved unsatisfactory performance with environmental dynamics. To address this issue, we propose GraphLoc, a novel approach to enhancing fingerprint-based localization with graph representation learning which can encode the structural information underlying radio fingerprints for robust localization. We first adopt graph signal processing of CSI fingerprints to create an unweighted graph. GraphLoc transforms the tasks of location estimations into node classification in a constructed graph. Then, we develop multilayer graph attention networks (GATs) with the residual structure (Res-GAT) to learn graph representation by collecting the neighboring node features and aggregating their neighboring embeddings. Furthermore, in order to guarantee and speed up our Res-GAT convergence, we propose a training strategy to overcome training difficulty and overfitting for improving the quality of graph representation. Finally, extensive experimental results in many typical indoor scenarios demonstrate that the GraphLoc system can achieve better accuracy than other comparative schemes, even with the robustness of environmental dynamics, effectively facilitating fingerprint-based localization for fully practical LBS.