A Gravity-informed Spatiotemporal Transformer for Human Activity Intensity Prediction | JoVE Visualize

Area of Science:

Artificial Intelligence
Geospatial Science
Physics

Background:

Human activity intensity prediction is vital for location-based services.
Existing models often neglect physical spatial interaction constraints, causing issues like over-smoothing and uninterpretable correlations.
There is a need for more interpretable and physically grounded spatiotemporal prediction models.

Purpose of the Study:

To propose a novel physics-informed deep learning framework, Gravityformer, for human activity intensity prediction.
To integrate the universal law of gravitation into transformer attention mechanisms to refine spatial interaction modeling.
To address limitations of existing methods, including uninterpretable spatial correlations and the over-smoothing phenomenon.

Main Methods:

Developed a physics-informed deep learning framework, Gravityformer, integrating the universal law of gravitation.
Estimated spatially explicit mass parameters using spatiotemporal embedding features.
Modeled spatial interaction via an adaptive gravity model within an end-to-end neural network.
Utilized learned spatial interactions to mitigate over-smoothing in transformer attention.
Proposed a parallel spatiotemporal graph convolution transformer for balanced spatial-temporal learning.

Main Results:

Gravityformer demonstrated quantitative and qualitative superiority over state-of-the-art benchmarks across six real-world datasets.
The learned gravity attention matrix showed interpretability based on geographical laws.
The model improved generalization capabilities in zero-shot cross-region inference.
Successfully mitigated the over-smoothing phenomenon in spatiotemporal prediction.

Conclusions:

Integrating physical laws, specifically the law of gravitation, with deep learning offers a novel approach for spatiotemporal prediction.
Gravityformer provides a more interpretable and physically grounded method for human activity intensity prediction.
The framework enhances the accuracy and generalization of location-based services by refining spatial interaction modeling.