Context-aware DL-based Approach Combining GNN and Attention for CSI-based Indoor Localization

In recent years, due to the limitations of GPS in indoor environments, where signals may be obstructed or reflected, significant advancements have been made in indoor localization technologies to address this gap. Among these, fingerprinting localization based on Channel State Information (CSI) has garnered widespread attention owing to its robustness and cost-effectiveness. The inherent nature of deep learning (DL), with its offline training and online positioning capabilities, aligns seamlessly with CSI fingerprinting localization. With the emergence of deep learning techniques, fingerprinting localization has undergone transformative changes. However, most of the current DL-based approaches are inspired by computer vision and natural language processing, implying that they do not fully account for the physical characteristics of CSI. In this paper, we thoroughly consider the physical properties of CSI and design a context-aware approach to embed CSI into a non-Euclidean space for graph-structured learning. Moreover, in feature learning, we separately contemplate the spatial and frequency dimensions, engaging in graph-structured feature learning from distinct physical dimensions. Results on real-world datasets demonstrate that our methodology significantly outperforms state-of-the-art DL-based and meachine learning (ML)-based approaches.