<p>Cellular-based localization has emerged as a promising alternative for mobile device positioning during interruptions or degradation of the Global Navigation Satellite System (GNSS), especially in urban environments where accurate positioning remains important for navigation, mobility monitoring, and location-based services. Various approaches have been proposed, ranging from simple proximity-based methods to more advanced machine learning techniques. Motivated by the observation that the input to this problem—cell identification, cell location, and the Received Signal Strength Indicator (RSSI) measured at the mobile device—can be naturally represented as graph-structured data, this paper proposes a Graph Neural Network (GNN)-based method for cellular localization. By modeling Base Transceiver Stations (BTSs), users, and their interactions as nodes and edges in a graph, the proposed framework captures both local and global spatial dependencies to improve positioning accuracy. Comprehensive experiments were conducted on two independent real-world datasets under three scenarios representing different spatial and temporal conditions. In all scenarios, the proposed GNN-based method consistently achieved the best performance compared with classical, fingerprinting, and Convolutional Neural Network (CNN)-based approaches. It attained mean localization errors of 16.6 m, 31.0 m, and 54.6 m in Scenarios 1, 2, and 3, respectively, showing consistent performance across different spatial and temporal conditions. In contrast, fingerprint-based methods were not applicable in environments where corresponding fingerprints were unavailable.</p>