TrajFlow: Nation-wide Pseudo GPS Trajectory Generation with Flow Matching Models

TrajFlow is the first flow-matching-based generative model that overcomes the scalability, diversity, and efficiency limitations of existing diffusion-based approaches to produce high-fidelity, nation-wide pseudo-GPS trajectories for applications in urban planning and traffic management.

The Gist

Imagine you are an urban planner trying to design a new subway line or prepare for a natural disaster. You need to know how millions of people move around a country every day. But you can't just ask everyone to share their private GPS data from their phones; that would be a massive privacy nightmare.

So, scientists try to build "fake" GPS data that looks and acts exactly like the real thing, without revealing any real people's secrets. This is called Pseudo-GPS Trajectory Generation.

For a long time, the best tools for this job were like slow, meticulous painters. They would start with a blank canvas (random noise) and slowly, step-by-step, add details until a picture of a person's journey appeared. This worked okay for small cities, but when they tried to paint a picture of an entire country, the process became incredibly slow, and the details got blurry or messy.

Enter TrajFlow, the new model introduced in this paper. Think of TrajFlow not as a painter, but as a high-speed, smart river guide.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Zoom" Issue

Imagine you are looking at a map.

• Zoomed in (City level): You see a person walking down a street. The details are clear.
• Zoomed out (Country level): You see a person flying from Tokyo to Osaka.

Old models (called "Diffusion models") struggled when you tried to do both at once. It's like trying to use the same brush size to paint a tiny ant and a giant whale. If you use a brush big enough for the whale, you miss the ant's details. If you use a small brush for the ant, painting the whale takes forever. This made it hard to generate realistic travel data for an entire nation.

2. The Solution: The "Flow" Approach

TrajFlow uses a new technique called Flow Matching.

• The Old Way (Diffusion): Imagine trying to un-mix a cup of coffee and milk by slowly stirring it backward. It takes many, many slow steps to get the milk and coffee separated perfectly.
• The New Way (Flow Matching): Imagine the coffee and milk are connected by a straight, invisible river. TrajFlow learns the direction of that river. Instead of taking 100 slow steps to un-mix the coffee, it just follows the river current directly from the "mixed" state to the "separated" state. It's much faster and smoother.

3. The Secret Sauce: "Harmonization" (The Translator)

Even with the fast river, there was still a problem: The data was too messy. A short walk in a neighborhood and a long train ride across the country are measured in very different scales.

• The Analogy: Imagine trying to teach a robot to draw both a tiny ant and a giant elephant. If you just say "draw it," the robot gets confused by the size difference.
• TrajFlow's Trick: Before the robot draws, it uses a translator. It shrinks the giant elephant down to the size of the ant and stretches the ant up to the size of the elephant, putting them both on the same "standardized" canvas. The robot learns the shape of the journey on this standard canvas. Once the drawing is done, the translator zooms it back out to the real-world size. This ensures the model doesn't get confused by the huge differences in distance.

4. Why This Matters

The researchers tested TrajFlow using millions of real GPS records from people all over Japan.

• Speed: It generates data 10 times faster than the old methods because it takes fewer steps (like taking a direct train instead of a bus with 30 stops).
• Accuracy: It works great whether you are looking at a single neighborhood, a whole city like Tokyo, or the entire country of Japan.
• Variety: It doesn't just copy-paste routes. It learns the rules of travel (like "people take trains for long distances and walk for short ones") and creates brand new, realistic journeys that have never existed before.

The Big Picture

TrajFlow is like a "Mobility Simulator" for the future.

Because it can generate millions of fake but realistic travel paths for an entire country in minutes, city planners, disaster responders, and traffic engineers can use it to:

• Test how a new bridge would affect traffic.
• Simulate how people would evacuate during an earthquake.
• Plan better public transport routes without ever needing to invade anyone's privacy.

In short, TrajFlow takes the messy, slow, and privacy-heavy job of understanding human movement and turns it into a fast, clean, and safe process that can handle the whole country at once.

Technical Summary

1. Problem Definition

The paper addresses the critical need for generating synthetic GPS trajectory data to overcome limitations associated with real-world data, such as privacy concerns, restricted accessibility, and high acquisition costs. While existing generative models (particularly diffusion-based ones like DiffTraj) have shown promise, they face three major bottlenecks when scaling from urban to nationwide levels:

• Multi-scale Capability: Existing models struggle to generalize across different spatial scales (e.g., from city blocks to entire nations). As the scale increases, the signal-to-noise ratio (SNR) drops significantly, causing diffusion models to fail in reconstructing fine-grained local structures amidst large-scale variations.
• Transportation-Mode Diversity: Most current approaches focus heavily on taxi data, failing to capture the diversity of real human mobility modes (walking, cycling, driving, public transit).
• Efficiency and Robustness: Diffusion models require iterative denoising steps (often hundreds) to generate high-quality data, making them computationally expensive. Furthermore, their fixed-magnitude noise injection is ill-suited for datasets spanning vastly different numerical ranges (e.g., short local trips vs. long-distance national travel).

2. Methodology: TrajFlow

The authors propose TrajFlow, the first flow-matching-based generative framework specifically designed for GPS trajectory generation. The methodology integrates three core components:

A. Flow Matching Paradigm

Instead of using the stochastic reverse-time SDEs of diffusion models, TrajFlow employs Conditional Flow Matching (CFM).

• Mechanism: It learns a time-dependent vector field $v_\theta(x, t)$ that transforms a simple prior distribution (Gaussian noise) into the complex target data distribution via an Ordinary Differential Equation (ODE).
• Advantage: This approach allows for direct regression on the target vector field along a prescribed probability path, offering greater training stability and robustness across multi-scale scenarios compared to diffusion objectives.

B. Trajectory Harmonization & Reconstruction

To address the SNR imbalance and computational overhead, TrajFlow introduces a preprocessing and postprocessing pipeline:

• Normalization: Raw GPS coordinates are normalized into a shared bounded space to prevent large-scale variations from overwhelming small-scale local displacements.
• Feature Transformation (RDP): The Ramer-Douglas-Peucker (RDP) algorithm is used to compress raw trajectories (e.g., ~120 points) into a compact set of key waypoints (e.g., ~10 points). This preserves the essential geometric shape while drastically reducing sequence length, improving training efficiency and stability.
• Reconstruction: After generation in the normalized, compressed space, the model inverts the normalization and interpolates the points back to real geographic coordinates.

C. Model Architecture

• Conditioning: The model uses a Wide & Deep architecture to encode heterogeneous inputs, including departure time, Origin-Destination (OD) zones, and transportation modes.
  • Wide Component: Linear projection of numeric features (speed, distance).
  • Deep Component: Embeddings for discrete features (time bins, OD IDs, mode).
• Backbone: A U-Net architecture parameterizes the vector field, with condition embeddings injected into every block via additive bias.
• Training Objective: The model minimizes a regression loss between the predicted vector field and the ground-truth vector field defined by a straight path between noise and data. An auxiliary supervised loss is added to enhance spatial awareness of fine OD locations.

3. Key Contributions

1. Novel Paradigm: TrajFlow is the first application of flow matching to GPS trajectory generation, demonstrating superior robustness and stability across multi-scale scenarios compared to diffusion baselines.
2. Methodological Innovation: The integration of trajectory harmonization (RDP compression) and OD-conditioned normalization creates a unified framework that simultaneously addresses scalability, diversity, and efficiency.
3. First Nationwide Generator: It is the first model capable of generating high-fidelity pseudo-GPS trajectories at the nationwide scale (Japan), moving beyond the limited urban/metropolitan scopes of previous works.

4. Experimental Results

The model was evaluated on a massive, private dataset containing millions of trajectories across Japan (covering Central Tokyo, Tokyo Metropolis, and Nationwide scales).

• Performance Metrics: Evaluated using Density JS Divergence (spatial distribution), Dynamic Time Warping (DTW), and Fréchet distance (geometric shape).
• Comparative Advantage:
  • Nationwide Scale: TrajFlow significantly outperformed DiffTraj, TrajGAN, and TrajVAE. For instance, in the nationwide setting, TrajFlow achieved a median DTW of 10.977 km compared to DiffTraj's 451.042 km, and a much lower Density JS (0.227 vs. 0.673).
  • Multi-scale Robustness: While diffusion models degraded sharply as spatial scale increased, TrajFlow maintained consistent performance from urban to nationwide levels.
  • Transportation Diversity: The model successfully preserved the distinct trip-length distributions for different modes (train, car, bike, walk), matching ground-truth heterogeneity.
• Efficiency: TrajFlow achieves high fidelity with only ~10 ODE integration steps. In contrast, diffusion baselines required 200–300 steps to approach similar (but still inferior) performance, making TrajFlow significantly faster in both training and inference.

5. Significance

• Urban Planning & Policy: TrajFlow enables the creation of large-scale, privacy-preserving mobility datasets essential for inter-region urban planning, traffic management, and disaster response simulation without compromising individual privacy.
• Scientific Advancement: It establishes Flow Matching as a superior alternative to Diffusion Models for high-dimensional, multi-scale geospatial data generation, solving the "scale mismatch" problem inherent in current SOTA methods.
• Privacy Preservation: By generating pseudo-data that captures aggregate patterns rather than individual trajectories, the model facilitates research and development in mobility systems while adhering to strict privacy constraints.

In conclusion, TrajFlow represents a significant leap forward in synthetic mobility data generation, offering a scalable, efficient, and high-fidelity solution for modeling human movement across entire nations.