A Deep Gravity model for mobility flows generation.

Area of Science:

• Computational Social Science
• Geographic Information Science
• Artificial Intelligence

Background:

• Human mobility patterns are crucial for understanding societal dynamics, including epidemic spread and environmental impact.
• Accurate modeling of mobility flows is essential, especially when direct data is unavailable.
• Existing models often struggle with the complexity and non-linear relationships inherent in mobility data.

Purpose of the Study:

• To introduce Deep Gravity, a novel deep learning model for generating human mobility flow probabilities.
• To leverage diverse geographic features and voluntary geographic data for enhanced mobility prediction.
• To evaluate Deep Gravity's performance against established models and assess its generalization capabilities.

Main Methods:

• Development of the Deep Gravity model utilizing deep neural networks.
• Integration of various geographic features such as land use, road networks, and facility locations.
• Experimental validation using mobility flow data from England, Italy, and New York State.

Main Results:

• Deep Gravity demonstrated significantly improved performance in predicting mobility flows compared to traditional gravity models.
• The model showed particular effectiveness in densely populated regions.
• Deep Gravity exhibited strong generalization, accurately predicting flows in areas without prior training data.

Conclusions:

• Deep Gravity offers a powerful and effective approach for modeling human mobility flows, outperforming existing methods.
• The model's ability to generalize and its explainability using AI techniques provide valuable insights into geographic influences on mobility.
• This research advances the understanding and prediction of mobility patterns with significant societal implications.