Federated Learning for Privacy-Aware Human Mobility Modeling.

Area of Science:

• Artificial Intelligence
• Computer Science
• Data Science

Background:

• Human mobility modeling is crucial for spatiotemporal event prediction (e.g., traffic, disease spread).
• Current deep learning models require extensive, privacy-sensitive spatiotemporal data.
• Federated Learning (FL) offers a privacy-preserving alternative by avoiding data centralization.

Purpose of the Study:

• To investigate the efficacy of Federated Learning for human mobility modeling.
• To compare FL implementations of state-of-the-art models (Flashback, DeepMove) against centralized approaches.
• To analyze the impact of FL parameters on model performance and stability.

Main Methods:

• Implemented and compared centralized models: Gated Recurrent Unit (GRU), Flashback, and DeepMove.
• Developed Federated Learning versions of the best-performing centralized models (Flashback and GRU).
• Evaluated models on large-scale Foursquare and Gowalla datasets, analyzing performance metrics and training stability.

Main Results:

• Federated Learning versions of Flashback and GRU demonstrated less stable training with higher loss variability.
• FL models exhibited slower convergence and reduced performance compared to their centralized counterparts.
• Model performance was significantly affected by the number of clients and data sparsity.

Conclusions:

• Federated Learning presents technical challenges for state-of-the-art deep learning in human mobility.
• Privacy-preserving FL models may require further optimization to match centralized model performance.
• Future research should address FL stability and parameter tuning for complex spatiotemporal tasks.