Demonstrating Deep Learning-based Spatial Diffusion

Abstract

Metadata geolocation, i.e., mapping information collected at a cellular Base Station (BS) to the geographical area it covers, is a central operation in producing statistics from mobile network measurements. This task requires modeling the probability that a device attached to a BS is at a specific location, and it is currently accomplished via simplistic approximations based on Voronoi tessellations. However, Voronoi cells exhibit poor accuracy compared to real-world geolocation data, which can reduce the reliability of downstream research pipelines. To overcome this limitation, DEEPMEND proposes a new data-driven approach relying on a teacher-student paradigm that combines probabilistic inference and deep learning. Similarly to other benchmarks, DEEPMEND can produce geolocation maps using only the BS positions, yielding a 56% accuracy gain compared to Voronoi tessellations. Our demonstrator will show visual and qualitative comparisons between DEEPMEND and several competitor approaches, allowing users to explore BS deployments from different geographical regions and operators.