Spatiotemporal-Augmented Graph Neural Networks for Human Mobility Simulation

TL;DR

STAR introduces a spatiotemporal-augmented graph framework to simulate human mobility by capturing dynamic location effects via multi-channel graphs (SDG, TTG, STG), a dual-branch dwell-aware decision generator, and an adversarial discriminator trained with policy gradients. The approach explicitly models spatiotemporal correspondences and location dwell durations, addressing exposure bias and long-term trajectory patterns. Across four real datasets, STAR outperforms state-of-the-art baselines on multiple fidelity and utility metrics, with ablation studies confirming the value of each graph channel and the dwell branch. The work offers a scalable, graph-based solution for realistic mobility simulation with potential applications in policy testing and urban planning, while noting limitations in dynamic scenarios and data scarcity and proposing future extensions leveraging external factors and cross-dataset mobility commonalities.

Abstract

Human mobility patterns have shown significant applications in policy-decision scenarios and economic behavior researches. The human mobility simulation task aims to generate human mobility trajectories given a small set of trajectory data, which have aroused much concern due to the scarcity and sparsity of human mobility data. Existing methods mostly rely on the static relationships of locations, while largely neglect the dynamic spatiotemporal effects of locations. On the one hand, spatiotemporal correspondences of visit distributions reveal the spatial proximity and the functionality similarity of locations. On the other hand, the varying durations in different locations hinder the iterative generation process of the mobility trajectory. Therefore, we propose a novel framework to model the dynamic spatiotemporal effects of locations, namely SpatioTemporal-Augmented gRaph neural networks (STAR). The STAR framework designs various spatiotemporal graphs to capture the spatiotemporal correspondences and builds a novel dwell branch to simulate the varying durations in locations, which is finally optimized in an adversarial manner. The comprehensive experiments over four real datasets for the human mobility simulation have verified the superiority of STAR to state-of-the-art methods. Our code is available at https://github.com/Star607/STAR-TKDE.

Figures (5)

• Figure 1: An illustration example. (a) depicts that various locations (HOME, Bar, Cafe, Pizza) get busy when people get off work from the office, revealing the spatiotemporal correspondences among locations. (b) depicts the successive visits of locations in an individual trajectory, resulting in varying durations in different locations.
• Figure 2: The overall framework of STAR. Firstly, given an observed human trajectory, the multi-channel embedding module generates location embeddings based on the proposed multi-channel spatiotemporal graphs. Secondly, the decision generator module predicts the future trajectory by balancing the exploration branch which is prone to another location and the dwell branch which decides whether to stay at the previous location. Finally, STAR is optimized in an adversarial manner by the policy discriminator module to alleviate the exposure bias of the maximum likelihood manner.
• Figure 3: Ablation study on the channels of graphs. All experimental results are conducted over five trials for a fair comparison. STG, SDG and TTG represent SpatioTemporal Graph, Spatial Distance Graph and Temporal Transition Graph respectively. MSC and SGP are short for Moscow and Singapore. A lower JSD value indicates a better performance.
• Figure 4: Effects of the number of layers and attention heads on STAR. All experimental results are conducted over five trials.
• Figure 5: The geographical visualization of location visit frequency of real and simulated human trajectories on the NYC dataset.