ARPIA: A multi-scale heterogeneous graph neural network for mobile device geolocation with explicit non-connection modeling

This work presents ARPIA (Advanced Radio Positioning Intelligence Architecture), a multi-scale heterogeneous graph neural network for mobile device localization that explicitly models non-connections between devices and base stations through a physics-informed force system. Experimental validation using a dataset of 38,356 devices (47,619 total nodes) from the Campinas metropolitan region demonstrates robust generalization capability. On the independent test set ( N = 5 , 754 ), ARPIA achieves a mean absolute error of 34.6 m (92.9% accuracy within 100 m), while maintaining a fitting error of 7.55 m (100% accuracy) on the training distribution, confirming the architecture’s capacity to capture complex geometric dependencies. Ablation studies reveal that removing physics-informed constraints or local precision modules degrades learning capacity by 9-12 × , confirming their critical importance. ARPIA significantly outperforms classical methods (RSSI, TOA, TDOA) and achieves 1.67 × improvement over the best alternative deep learning configuration (GATv2, HAN, RGCN), demonstrating the effectiveness of explicit non-connection modeling for wireless network localization.