Fine Grained Urban Grids Clustering of Mobile Phone Metadata with Deep Spatial Temporal Clustering

Mobile phone metadata is widely used to extract socio-economic activity metrics at scale. Properly utilized, this data provides unique insights into downstream tasks and business value. This paper uses large scale multi-region service-level mobile data traffic to investigates geospatial temporal clustering, aiming to cluster similar urban grids sharing the same social signature. Leveraging deep learning, time series clustering transcends morphology, achieving more accurate clustering in the hidden space. However, most current algorithms struggle to extract both temporal and spatial information. We propose a novel method, Deep Spatial Temporal Clustering (DSTC), for clustering fine-grained urban grids based on geospatial temporal characteristics of mobile traffic data. DSTC employs a lightweight spatiotemporal autoencoder and graph construction, optimizing clustering and reconstruction objectives to learn representations and cluster assignments. Using data from the NetMob 2023 Data Challenge, we demonstrate DSTC’s feasibility and interpretability through extensive case studies on two French metropolitan cities and three applications. This work advances geospatial temporal clustering and offers practical insights and business value from mobile phone metadata, reflecting real-world industrial scenarios. The source code for this paper is available at https://github.com/ComplexNetTSP/DSTCourGitHubrepository