Sensing overlapping geospatial communities from human movements using graph affiliation generation models

Geographical units densely connected by human movements can be treated as a geospatial community. Detecting geospatial communities in a mobility network reveals key characteristics of human movements and urban structures. Recent studies have found communities can be overlapping in that one location may belong to multiple communities, posing great challenges to classic disjoint community detection methods that only identify single-affiliation relationships. In this work, we propose a Geospatial Overlapping Community Detection (GOCD) framework based on graph generation models and graph-based deep learning. GOCD aims to detect geographically overlapped communities regarding the multiplex connections underlying human movements, including weak and long-range ties. The detection process is formalized as deriving the optimized probability distribution of geographic units' community affiliations in order to generate the spatial network, i.e., the most reasonable community affiliation matrix given the observed network structure. Further, a graph convolutional network (GCN) is introduced to approach the affiliation probabilities via a deep learning strategy. The GOCD framework outperformed existing baselines on non-spatial benchmark datasets in terms of accuracy and speed. A case study of mobile positioning data in the Twin Cities Metropolitan Area (TCMA), Minnesota, was presented to validate our model on real-world human mobility networks. Our empirical results unveiled the overlapping spatial structures of communities, the overlapping intensity for each CBG, and the spatial heterogeneous structure of community affiliations in the Twin Cities.