A Graph-Based Framework for Analysing Human Mobility Using Spatial-Temporal Network Representations

Predicting human mobility requires choices about how to represent people, places, and time. This paper provides a graph-construction and task-formulation framework for mobility analysis using passively collected trajectories. Using the Humob 2023 dataset, we instantiate three commonly used graph representations then use Graph Attention Networks (GAT) and Graph Convolutional Network (GCN) baselines to separate representation effects from architectural effects. The results indicate that representation and task definition dominate performance differences, while the gap between GAT and GCN is comparatively smaller in this setting. We distil practical considerations for selecting mobility graph representations and discuss limitations, especially the need for explicit dynamic-graph or sequence modelling when fine-grained temporal evolution is central.