Towards Cooperative Localization With Implicit Connectivity: Graph Neural Network Approach

Abstract

This paper aims to facilitate graph neural network (GNN)-based cooperative localization (CL) in scenarios where the connections between gNodeBs (gNBs) and those between user equipment (UEs) are not given, which are yet crucial for determining cooperation pairs. To address this, we first define the concept of implicit connectivity where UE-UE and gNB-gNB connections are conjectured from the available explicit gNB-UE connectivity such that implicit connection between UEs (or gNBs) is likely to exist when there are shared gNBs (or UEs). Considering implicit connections along with explicit ones makes the graph denser, helping spread valuable information throughout the network. Besides, the numbers of shared gNBs and UEs are utilized to enhance node features through a self-attention-based feature embedding, which is beneficial for training the subsequent GNN. Through a realistic dataset generated by a ray-tracing simulator, we verify that the proposed technique achieves the 90th percentile error of 1.5031 (m), significantly outperforming all benchmarks and satisfying the 6G positioning requirement.