In an ideal world, trajectory data would be a continuous, high-frequency stream of coordinates—a perfect digital twin of movement. In reality, spatiotemporal data is "holey." Whether due to battery-saving protocols on mobile devices, signal interference in urban canyons, or privacy-preserving downsampling, the data we actually work with is often a collection of fragmented, sparse observations.
The challenge in modern mobility research is not just to analyze these fragments, but to intelligently "stitch" them back into a continuous whole. This involves moving from simple interpolation to sophisticated generative and geometric models that understand the underlying physics and logic of movement.
The Connectivity Problem
Traditional models for next-location prediction or traffic simulation often rely on the "stay-on-the-data" dogma. However, as Jianxiong Li and Xianyuan Zhan argue in their work on DOGE, real-world datasets contain significant "holes" along critical paths. If an algorithm cannot generalize across these gaps, it fails to synthesize a complete solution.
The problem is twofold:
- Temporal Sparsity: Large time intervals between pings make it unclear which route was taken.
- Spatial Sparsity: Many geographic areas have little to no historical data, creating "cold start" problems for routing and recommendation.
Reconstructing Continuity via Diffusion and ODEs
To fill these gaps, researchers have moved toward generative frameworks that treat trajectory recovery as a denoising or flow problem.
Jinming Wang and Hai Wang introduced TrajWeaver, which uses a State Propagation Diffusion Model (SPDM). Instead of treating each missing point as an independent guess, the model allows information to flow between denoising steps. This ensures that the reconstructed path is not just a set of likely points, but a physically plausible sequence where each step logically follows the last.
Similarly, Jinsung Jeon and Noseong Park utilize Neural Ordinary Differential Equations (NODEs) in their LightMove framework. By modeling movement as a continuous-time function rather than a discrete sequence, NODEs can naturally handle irregular sampling intervals, "filling in" the state of a vehicle at any arbitrary timestamp.
Stitching Disjoint Trajectories
Sometimes the goal isn't just to fill a gap in one trip, but to connect segments from different users or sessions to understand a broader pattern. This is often referred to as "trajectory stitching."
In the context of offline reinforcement learning for medical treatments, the TreatStitch framework uses a Schrödinger Bridge. This mathematical tool from optimal transport finds the most likely path between two disparate states. By "stitching" together segments of real patient trajectories that don't naturally overlap, the model creates new, clinically plausible training sequences that help agents learn from sparse clinical data.
In urban mobility, Lei Zhang and Qingfu Fan apply a similar logic using Matrix Factorization in ESTP-MF. They treat the vast space of possible transitions as a sparse matrix and use factorization to "fill in the blanks" of unobserved but valid transitions, ensuring that a prediction model doesn't assign zero probability to a route simply because it hasn't been seen before.
Learning from Fragmented Signatures
When data is too sparse for exact reconstruction, researchers focus on "signatures"—robust representations that survive data loss.
Samta Shukla and colleagues proposed Spatio-Temporal Signatures, which represent messy GPS data as directed graphs. The intuition here is that even if 90% of the data points are missing, the general "flow" through spatial regions remains a powerful fingerprint for identifying similar users or behaviors.
This shift from point-based accuracy to topological consistency is a recurring theme. Whether it is Hua Wei integrating interpolation directly into imitation learning (ImIn-GAIL) or Tonglong Wei repurposing language models to "read" trajectories (PTR), the field is converging on a single realization: the "holes" in our data are not just noise to be removed, but gaps that can be bridged by modeling the underlying logic of how things move through space and time.