FedVoting: A Cross-Silo Boosting Tree Construction Method for Privacy-Preserving Long-Term Human Mobility Prediction.

Area of Science:

• Computational Social Science
• Machine Learning
• Data Privacy

Background:

• Human mobility prediction is crucial for urban planning and commercial applications.
• Privacy concerns limit the collection and sharing of personal GPS data, creating data silos.
• Traditional federated learning (FL) struggles with the complexity of long-term mobility data.

Purpose of the Study:

• To develop an optimized federated learning approach for accurate long-term human mobility prediction.
• To propose a novel federated voting (FedVoting) mechanism that ensures data privacy.
• To address the limitations of existing FL methods in handling irregular and complex mobility patterns.

Main Methods:

• Utilized a high-efficient gradient-boosting decision tree (GBDT) model within the FL framework.
• Introduced the FedVoting mechanism, aggregating differential privacy (DP)-protected GBDTs.
• Employed multiple training, cross-validation, and voting processes for optimal model generation.

Main Results:

• Achieved high accuracy in predicting long-term human mobility, including special event attendance and point-of-interest visits.
• FedVoting demonstrated significant accuracy improvements compared to independent silo training and state-of-the-art baselines.
• Performance was comparable to centralized training with only a negligible privacy cost.

Conclusions:

• The FedVoting mechanism effectively enhances long-term human mobility prediction accuracy.
• This approach successfully balances high performance with robust differential privacy protection.
• FedVoting offers a practical solution for leveraging distributed mobility data without compromising privacy.