Discovering Latent Spatial Invariance of Urban Wireless Data using Compression and Deep Learning

Abstract

Increasingly available high resolution geospatial wireless demand data is available from high density base stations, wireless localisation, and geo-tagged social media posts. Mapping the evolving spatiotemporal demand is critical for a wide range of infrastructure services, including future network planning and operations. However, monitoring geospatial data demand across a whole city is computationally and financially expensive. Here, we show that geospatial traffic demand data from both 0.4 million Twitter posts and 3.2 million base stations records can be compressed to spatially invariant points in London. These points correspond to major sources of human movement activity that act as either facilitators (e.g. public multi-modal transport hubs) or drivers (e.g. tourist attractions and business hubs). This demonstrates that by monitoring these spatially invariant critical points, we can obtain an accurate understanding of the human demand dynamics elsewhere in the city. Indeed, the operator which maps the dynamics between these points uncover the latent human connected dynamics embedded in complex urban ecosystems.We use both the latest signal processing technique of Graph Fourier Transform (GFT) and a AutoDecoder inspired deep learning neural network to demonstrate spatially invariant compression and both error-free and noisy recovery. These promising results show that we can exploit the connected structure of complex cities to dramatically reduce data monitoring.