Exploring Transformer and Graph Convolutional Networks for Human Mobility Modeling.

Area of Science:

• Artificial Intelligence
• Human Mobility Studies
• Data Science

Background:

• Accurate human mobility pattern estimation is crucial for urban planning, pollution control, and disease management.
• Next-place predictors, essential for mobility estimation, have not yet leveraged advanced AI techniques like GPT and GCNs.

Purpose of the Study:

• To explore the efficacy of General Purpose Transformer (GPT) and Graph Convolutional Network (GCN) based models for next-place prediction.
• To compare the performance of GPT and GCN models against a state-of-the-art model (Flashback-LSTM) on both sparse and dense mobility datasets.

Main Methods:

• Developed GPT- and GCN-based models adapted from general time series forecasting architectures.
• Evaluated model performance using two sparse (check-in) and one dense (continuous GPS) human mobility datasets.

Main Results:

• GPT-based models demonstrated a slight accuracy advantage over GCN-based models (1.0–3.2 p.p.) on tested datasets.
• Flashback-LSTM slightly outperformed GPT and GCN models on sparse datasets (1.0–3.5 p.p. difference).
• All evaluated models performed comparably on the dense GPS dataset.

Conclusions:

• GPT and GCN models show comparable performance to state-of-the-art methods in next-place prediction.
• The relevance of sparse data performance may diminish as future mobility data becomes denser (e.g., from smartphones).
• There is significant potential for GPT and GCN models to advance human mobility prediction beyond current state-of-the-art capabilities.