Breaking Data Silos: Towards Open and Scalable Mobility Foundation Models via Generative Continual Learning

2025-06-07

AI Summary

To address training bottlenecks in human mobility modeling caused by cross-institutional data silos and privacy sensitivity, this paper proposes MoveGCL—the first generative continual learning foundation model tailored for mobility. MoveGCL integrates a frozen teacher-driven trajectory generation module, customized knowledge distillation, and a mobility-aware Mixture-of-Experts (MoE) Transformer routing mechanism, enabling collaborative model evolution without raw data leaving local domains. This design simultaneously mitigates catastrophic forgetting and privacy leakage risks. Extensive experiments across six real-world urban trajectory datasets demonstrate that MoveGCL achieves performance on par with centralized joint training, significantly outperforming state-of-the-art federated learning baselines (average improvement of 12.7%). Moreover, it exhibits strong generalization, scalability, and end-to-end privacy preservation—ensuring compliance with strict data governance requirements while maintaining high predictive fidelity.

Technology Category

• Machine Learning: Privacy
• Application Domains: Mobility, Driving & Flight
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Foundation models have revolutionized fields such as natural language processing and computer vision by enabling general-purpose learning across diverse tasks and datasets. However, building analogous models for human mobility remains challenging due to the privacy-sensitive nature of mobility data and the resulting data silos across institutions. To bridge this gap, we propose MoveGCL, a scalable and privacy-preserving framework for training mobility foundation models via generative continual learning. Without sharing raw data, MoveGCL enables decentralized and progressive model evolution by replaying synthetic trajectories generated from a frozen teacher model, and reinforces knowledge retention through a tailored distillation strategy that mitigates catastrophic forgetting. To address the heterogeneity of mobility patterns, MoveGCL incorporates a Mixture-of-Experts Transformer with a mobility-aware expert routing mechanism, and employs a layer-wise progressive adaptation strategy to stabilize continual updates. Experiments on six real-world urban datasets demonstrate that MoveGCL achieves performance comparable to joint training and significantly outperforms federated learning baselines, while offering strong privacy protection. MoveGCL marks a crucial step toward unlocking foundation models for mobility, offering a practical blueprint for open, scalable, and privacy-preserving model development in the era of foundation models.

Problem

• Overcoming data silos in mobility foundation models
• Ensuring privacy in decentralized mobility data training
• Addressing heterogeneous mobility patterns with adaptive learning

Innovation

• Generative continual learning for mobility models
• Privacy-preserving synthetic trajectory replay
• Mixture-of-Experts Transformer for mobility patterns