Adapting to Open-World Events via Learning Spatio-Temporal Dynamics on Mobility Networks

Research Background

In modern society, the Mobility-as-a-Service (MaaS) system is seamlessly integrated by various transportation modes (such as public transport, ride-sharing, and shared bicycles). To achieve efficient MaaS operation, modeling the spatio-temporal dynamics of multimodal mobility networks is essential. However, existing methods either implicitly handle the interactions between different transportation modes or assume that these interactions are invariant. Moreover, when open-world events (such as holidays, adverse weather, and pandemics) occur, collective human mobility behavior will deviate significantly from normalcy, making this modeling task even more challenging.

Paper Source

This paper is co-authored by Zhaonan Wang from the Department of Geography and Geographic Information Science, University of Illinois at Urbana-Champaign, Renhe Jiang, Xuan Song, and Ryosuke Shibasaki from the Center for Spatial Information Science, The University of Tokyo, and Hao Xue, Flora D. Salim from the School of Computer Science and Engineering, University of New South Wales. The paper has been accepted by the journal Artificial Intelligence and will be officially published in 2024.

Research Content and Innovations

Research Procedure

This research focuses on the task of Multimodal Mobility Nowcasting. The authors first designed a Heterogeneous Mobility Information Network (HMIN) to explicitly represent the interactions (i.e., “intermodality”) between different transportation modes. Specifically, HMIN consists of region nodes, modal nodes, and various edges connecting them, simultaneously capturing the spatial and modal dependencies.

Secondly, the authors proposed a novel Memory-Augmented Dynamic Filter Generator (MDFG). MDFG can dynamically generate parameters based on the input sequence features, enabling the model to respond appropriately to various scenarios, including unseen open-world events. MDFG comprises a learnable mobility prototype memory and a dynamic filter generation layer. The mobility prototype memory stores high-level mobility representations containing spatial, temporal, and multimodal knowledge. The dynamic filter generation layer queries similar prototypes from the memory based on the current input sequence features and generates sequence-relevant parameters accordingly.

Finally, the authors combined HMIN and MDFG to design an Event-Aware Spatio-Temporal Network (EAST-Net). EAST-Net can not only explicitly model the dynamic interactions between different transportation modes but also adaptively adjust parameters based on the input sequence features, thus exhibiting strong event awareness and adaptability.

Datasets

The paper evaluates on five real-world mobility datasets spanning different spatio-temporal scales and containing significant open-world events with societal impacts, such as the “Jonas” blizzard in New York and Washington D.C. (2016), Hurricane “Dorian” in Florida (2019), and the COVID-19 pandemic in the United States and Chicago (2019).

Experimental Results

Experimental results demonstrate that EAST-Net achieves significantly better performance on the multimodal mobility nowcasting task compared to existing mainstream methods, especially when open-world events occur, showcasing its event awareness and adaptability.

Furthermore, through a series of knowledge transfer experiments, the researchers found that EAST-Net exhibits certain generalization capabilities in both spatial and temporal domains. Specifically, for the COVID-19 pandemic dataset, the knowledge learned from Chicago can be relatively directly applied nationwide. In contrast, for the blizzard dataset, knowledge transfer between cities requires further training adaptation. These findings suggest that the nature of mobility patterns (travel mode vs. destination) may affect the transferability of knowledge in different scenarios.

Research Significance

This research has the following main implications:

1) The HMIN structure explicitly represents the interactions between different transportation modes, laying the foundation for solving the multimodal mobility coordination and optimization problems in MaaS.

2) The MDFG mechanism enables the model to dynamically adjust parameters according to specific scenarios, exhibiting strong event awareness and adaptability, providing a referenceable solution for AI in open-world environments.

3) The generalization experiments suggest that the proposed memory and prototype representation mechanism has the potential to become a core component of future spatio-temporal foundational models, enabling models to adapt not only to different spatial regions but also to respond promptly to emerging open-world events.

Summary

This research tackles the problem of modeling spatio-temporal dynamics of multimodal mobility networks, particularly in adapting to sudden anomalies caused by open-world events. The authors designed a Heterogeneous Mobility Information Network and a memory-based dynamic parameter generation mechanism, proposing an Event-Aware Spatio-Temporal Network model. Extensive experimental results validate the model’s superior performance in event awareness, adaptability, and generalization capability, providing valuable technical support for practical applications such as MaaS. Looking ahead, this field deserves continued attention and in-depth exploration from both industry and academia.