Introduction
The rise of Foundation Models (FMs) has revolutionized natural language processing and computer vision by leveraging massive, centralized datasets. However, applying this paradigm to human mobility—tracking how people move through cities via GPS, transit records, and check-ins—presents a unique set of challenges. Mobility data is inherently sensitive, governed by strict privacy regulations like GDPR, and often locked within corporate or municipal "data silos."
To bridge this gap, the research introduces MoveGCL, a framework designed to build mobility foundation models without requiring the centralized sharing of raw data. By utilizing generative continual learning, MoveGCL allows a model to learn from new cities sequentially while retaining knowledge of previous ones, effectively breaking the silos that hinder large-scale mobility AI.
The Challenge of Mobility Data Silos
Training a model that understands movement patterns across diverse urban environments (e.g., New York and Paris) typically forces researchers into three suboptimal paths:
- Isolated Training: Models are trained on single cities, leading to poor generalization across different urban layouts.
- Joint Training: All data is moved to a central server, creating significant privacy risks and regulatory hurdles.
- Federated Learning: While privacy-preserving, it requires high communication overhead and is difficult to scale across many institutions.
The core technical hurdle in sequential learning is catastrophic forgetting: when a model learns a new city (City B), it often loses the specialized patterns it learned from the previous one (City A). MoveGCL aims to solve this without ever needing to revisit City A's raw data.
Methodology: The MoveGCL Framework
MoveGCL addresses these issues through a multi-pronged architectural approach:
1. Generative Replay with a Teacher-Student Architecture
Instead of storing raw historical data, MoveGCL uses a "Teacher" model (the version trained on previous cities) to generate synthetic trajectories. When the model encounters a new city, the "Student" model learns from both the new real-world data and the "hallucinated" data from the Teacher. This synthetic replay anchors the model’s existing knowledge without ever needing to access the original sensitive datasets.
2. Mobility-Aware Mixture-of-Experts (MoE)
Human movement is not monolithic; a short walk to a coffee shop differs fundamentally from a long-distance commute. MoveGCL employs a Mixture-of-Experts architecture where specialized sub-networks handle different behavioral patterns. A dedicated router analyzes features like "jump distance" and "waiting time" to assign trajectories to the most appropriate expert, allowing the model to transfer "movement logic" across cities with different layouts.
3. Layer-wise Progressive Adaptation
To maintain stability during sequential learning, MoveGCL utilizes progressive adaptation. Rather than updating all parameters simultaneously, the framework unfreezes layers incrementally—from the outer layers inward. This ensures that new information is integrated without overwriting the foundational geographic knowledge stored in the deeper layers.
4. Unified Location Encoder
To avoid city-specific biases, MoveGCL uses a continuous representation for locations based on GPS coordinates, Point-of-Interest (POI) categories, and "mobility heat" (density). This creates a universal geographic language, allowing the model to function across different coordinate systems and urban structures.
Experimental Evaluation
The researchers evaluated MoveGCL using high-resolution mobility data from six major US cities: Atlanta, Chicago, Los Angeles, New York, Seattle, and Washington D.C. The combined dataset represented over 800,000 users and millions of individual trajectories.
MoveGCL was benchmarked against traditional Markov models, LSTMs, and state-of-the-art trajectory models like TrajBert and CLET, as well as privacy-preserving federated learning methods like PMF and LightTR.
Key Insights and Results
- Performance Parity: MoveGCL achieved performance levels comparable to "Joint Training"—the theoretical gold standard where all data is shared. This demonstrates that generative replay is a viable substitute for raw data access.
- Accuracy Improvements: The model showed an average 8% improvement in Acc@1 (predicting the exact next location) compared to standalone and federated baselines.
- Order Invariance: A critical finding was that the sequence of cities learned (e.g., Chicago then Seattle vs. Seattle then Chicago) did not significantly impact final performance, indicating a robust and stable learning process.
- Privacy Resilience: In Membership Inference Attacks, the success rate for attackers was near 50%, equivalent to a random guess. With a low privacy budget ($\epsilon \approx 1-2$), the model proved highly resistant to data leakage.
Limitations and Future Directions
While MoveGCL represents a significant step toward open mobility models, several areas remain for future exploration. The current study focused exclusively on US cities; mobility patterns in the Global South or rural regions may require further tuning. Additionally, while the generative approach provides a natural privacy buffer, integrating formal Differential Privacy (DP) guarantees into the training loop is a priority for future iterations. Finally, the researchers emphasize the need for strict governance to prevent the misuse of high-fidelity mobility models for unauthorized surveillance.
Conclusion
MoveGCL offers a scalable, privacy-preserving path toward Foundation Models for human mobility. By replacing data sharing with generative knowledge transfer, it enables the creation of powerful, generalized models that respect the boundaries of data silos.
Summary updated