VAMBC: A Variational Approach for Mobility Behavior Clustering

Abstract

Many domains including policymaking, urban design, and geospatial intelligence benefit from understanding people’s mobility behaviors (e.g., work commute, shopping), which can be achieved by clustering massive trajectories using the geo-context around the visiting locations (e.g., sequence of vectors, each describing the geographic environment near a visited location). However, existing clustering approaches on sequential data are not effective for clustering these context sequences based on the contexts’ transition patterns. They either rely on traditional pre-defined similarities for specific application requirements or utilize a two-phase autoencoder-based deep learning process, which is not robust to training variations. Thus, we propose a variational approach named VAMBC for clustering context sequences that simultaneously learns the self-supervision and cluster assignments in a single phase to infer moving behaviors from context transitions in trajectories. Our experiments show that VAMBC significantly outperforms the state-of-the-art approaches in robustness and accuracy of clustering mobility behaviors in trajectories.