Abstract
Location data is a double-edged sword, powering essential services while posing significant privacy risks. This survey reviews 38 state-of-the-art works applying Federated Learning (FL) to spatio-temporal (ST) mobility data. We explore how decentralized learning addresses the privacy-utility tradeoff and the unique challenges of ST data, including high sparsity and non-IID distributions.
Introduction
Location data is incredibly sensitive; research indicates that just four spatio-temporal points can uniquely identify 95% of individuals. Traditional centralized machine learning requires aggregating raw data, creating massive privacy vulnerabilities. Federated Learning (FL) mitigates this by keeping data on local devices and sharing only model updates. However, ST mobility data introduces unique complexities:
- Auto-correlation: Spatial and temporal proximity implies natural relationships (e.g., traffic flow).
- Non-IIDness: Human mobility patterns are highly individualistic.
- Data Sparsity: Limited user check-ins make local model training difficult.
Federated Learning Architectures
FL implementations generally fall into two categories: Cross-Device and Cross-Silo. Cross-device FL involves a massive number of mobile or edge devices, while Cross-Silo FL involves a smaller number of organizations (e.g., hospitals or municipalities) acting as data silos. Understanding these architectures is crucial for selecting appropriate datasets and privacy tools.
Application Domains and Methodologies
The literature is categorized into three primary domains: Trajectory Prediction, Traffic Flow Prediction, and Location-Based Recommendations. A significant contribution of this survey is the application of Shannon’s Entropy to quantify mobility data "messiness," defining Horizontal Entropy ($H_u$) for user predictability and Vertical Entropy ($H_l$) for commonality across locations. This allows for more rigorous benchmarking against datasets like Foursquare, Gowalla, and PeMS.
Key Insights
- Mandatory Personalization: Due to the non-IID nature of mobility, global models often underperform. Techniques like Meta-Learning and local fine-tuning are essential to adapt models to individual habits.
- GNN Challenges: In traffic prediction, Graph Neural Networks (GNNs) struggle because individual silos only observe a fraction of the total urban graph, complicating the aggregation process.
- Privacy Layers: FL alone is insufficient; "curious" servers can still infer locations from model weights. Additional protections like Differential Privacy (DP) and Secure Multi-Party Computation (SMPC) are necessary for robust security.
Performance and Results
Federated models demonstrate competitive performance against centralized baselines. For trajectory accuracy, models like STSAN and STLPF achieved nearly 99% Average Percentile Rank (APR) on Foursquare datasets. In traffic management, CNFGNN achieved RMSE scores (e.g., 3.82 on PEMS-BAY) that are highly competitive with centralized GNNs, proving that privacy-preserving traffic management is technically viable.
Limitations and Future Work
The authors identify several critical gaps in current research:
- Byzantine Resilience: There is a lack of research on handling malicious participants who might poison models.
- Communication Efficiency: Moving large GNN parameters is expensive; more work is needed on model pruning and quantization.
- Semantic Integration: Incorporating location context (e.g., "hospital" vs. "park") could improve personalization.
- Standardization: The field lacks a common ground for data splitting and evaluation, making fair comparisons difficult.
Summary updated