Urban mobility analysis—the study of how people and vehicles move through a city—is the backbone of modern urban planning. It drives everything from optimizing traffic lights to designing efficient public transit systems. However, researchers face a persistent "data desert." Real-world GPS trajectory data is often proprietary (owned by private ride-hailing giants), privacy-sensitive (containing digital breadcrumbs that can reveal personal routines), or low quality (riddled with noise and missing points).
Existing public datasets are either too small or lack the quality needed for advanced deep learning applications. To overcome these privacy and accessibility barriers, researchers have introduced SynMob, a novel framework for generating high-fidelity synthetic GPS trajectory datasets using diffusion models. By releasing two massive synthetic datasets, the authors demonstrate that synthetic trajectories can effectively replace real-world data for critical tasks like traffic demand prediction and travel time estimation.
The SynMob Methodology
The authors employ a Diffusion Model as their core trajectory synthesizer. While diffusion models are famous for generating photorealistic images, SynMob creatively repurposes them for spatial-temporal data.
- The Process: The model uses a two-step approach. First, a forward noising process iteratively adds Gaussian noise to real trajectories until they become unrecognizable. Then, a reverse denoising process learns to "undo" this noise, effectively learning the underlying distribution of how people move.
- Architecture: The framework utilizes a UNet architecture optimized specifically for the unique sequential structure of trajectory data.
- Guidance: To ensure the generated paths are realistic and adhere to actual road networks, they use classifier-free diffusion guidance. This helps balance the diversity of the generated paths with the high fidelity required to match real-world constraints.
- Normalization: Practical adjustments, like length normalization, are implemented to ensure the model trains efficiently across varying trip durations.
Datasets and Key Insights
The paper introduces two major synthetic benchmarks trained using data from the GAIA project (Didi Chuxing):
- SYN-CHENGDU: 1 million records (4.39 GB), sampled at 3-second intervals.
- SYN-XI’AN: 1 million records (4.66 GB), also at 3-second intervals.
Compared to existing datasets, SynMob is unique in checking all the critical boxes: standard GPS trajectory format, high availability, high data quality, and guaranteed privacy.
Unlike older anonymization methods that simply "blur" real data points—which can often be reverse-engineered—the diffusion model learns the statistical distribution of movement. The resulting trajectories are entirely new but behave exactly like the originals. Furthermore, SynMob successfully captures complex spatial-temporal synchronization, maintaining realistic "day-night cycles" and road-network constraints (e.g., recognizing that traffic moves slower at 5 PM than at 3 AM).
Key Results and Validation
The ultimate test of synthetic data is its utility in the real world. The authors validated the datasets through rigorous statistical testing and downstream applications, proving that models trained on synthetic data can achieve nearly the same accuracy as those trained on real data.
- Statistical Fidelity: Using the Kolmogorov-Smirnov (K-S) test, the synthetic data showed a >90% similarity to real data across key metrics like travel distance, travel time, speed, and relative distance.
- Traffic Demand Prediction (TDP): When training state-of-the-art models (like AGCRN and GWNet) on synthetic data, the performance difference compared to real data was remarkably low—often under 6%.
- Travel Time Estimation (TTE): In predicting trip durations, the synthetic data maintained high consistency, with difference ratios generally staying within a modest range of 4-10%.
Limitations and Future Work
Despite this breakthrough, the authors acknowledge a few hurdles. The synthesizer still requires high-quality raw data as a "seed" to learn from initially; it cannot generate a city's mobility patterns entirely from scratch. Additionally, diffusion models are computationally heavy to train compared to simpler generative models.
Future work will focus on reducing this computational overhead and refining the synthesizer to capture even more subtle, long-tail mobility behaviors that might be missed in the current iteration.
The trajectory synthesizer and datasets are publicly available on GitHub at Applied-Machine-Learning-Lab/SynMob.
Summary updated