Real-world mobility data is rarely the continuous stream of coordinates that idealized models assume. Instead, it is a collection of "digital breadcrumbs"—fragmented, irregularly sampled, and often missing large segments due to signal loss, battery-saving duty cycles, or privacy-preserving filters. This data sparsity creates a fundamental gap: we see where an object was at 2:00 PM and 2:30 PM, but the critical "connective tissue" of the path taken, the road segments traversed, and the semantic intent of the journey is lost.
To bridge these gaps, researchers have moved beyond simple linear interpolation toward methods that treat trajectory recovery as a problem of semantic infilling, structural constraint, and probabilistic bridging.
Trajectories as Spatiotemporal Language
One of the most powerful intuitions in recent work is treating a sequence of locations like a sentence in a natural language. Just as a language model can predict a missing word based on the surrounding context, spatiotemporal models can "infill" missing segments by learning the underlying grammar of human movement.
Alessandro Crivellari et al. demonstrated this with TraceBERT, adapting the BERT architecture to reconstruct spatial-temporal gaps in motion trajectories. By treating discrete location sequences as tokens, the model learns to capture non-repetitive movement patterns that traditional routine-based methods miss. Similarly, Shang-Ling Hsu et al. proposed TrajGPT, which frames controlled synthetic trajectory generation as a multitask transformer problem, significantly improving temporal accuracy by modeling the "when" and "where" of mobility as a joint generative task.
Leveraging Structural and Physical Constraints
While language models rely on semantic context, mobility is physically constrained by road networks and geography. Sparse data often results in "virtual edges"—direct jumps between points that do not exist on a map.
To address this, Silin Zhou et al. introduced GRLSTM, which uses a graph-based Residual LSTM to compute trajectory similarity. By treating road points as a knowledge graph, the model can account for the physical infrastructure even when GPS pings skip multiple intersections. For high-resolution recovery, Tian Sun et al. developed TedTrajRec, which combines dynamic-aware graph neural networks (GNNs) with Neural Ordinary Differential Equations (ODEs). This allows the model to handle the irregular time gaps inherent in sparse data by treating movement as a continuous-time dynamical system rather than a discrete sequence.
Probabilistic Bridges and Generative Infilling
When data is extremely sparse, there may be multiple plausible paths between two points. Rather than predicting a single "correct" path, some researchers use generative models to sample the distribution of likely movements.
Ella Tamir proposed the Iterative Smoothing Bridge, which uses Schrödinger Bridges to learn data-driven dynamics that satisfy sparse, noisy intermediate observations. In the realm of urban planning, Yuanshao Zhu et al. introduced SynMob, utilizing diffusion models to generate high-fidelity synthetic trajectories. These synthetic datasets act as a proxy for sparse real-world data, allowing for robust traffic demand prediction without compromising individual privacy.
Robust Representation Learning
Even if a trajectory cannot be fully reconstructed, it can still be useful if we can map it into a robust vector space. The goal here is to ensure that two sparse trajectories representing the same route have similar "signatures."
The RED framework, proposed by Silin Zhou et al., achieves this through a self-supervised dual-objective that masks road segments during training. This forces the model to learn the underlying spatial and temporal dependencies, making the resulting embeddings resilient to the gaps common in point-of-interest (POI) check-in data. For recommendation tasks, Qi Jiang utilized multi-granularity contrastive learning to align different semantic views of a user's movement, ensuring that sparse check-ins still yield accurate next-location predictions.
Efficiency in Scale
As datasets grow to include millions of vehicles, the computational cost of recovery becomes a bottleneck. Jieming Shi et al. addressed this with TRMMA, a dual-transformer architecture that treats map matching as a localized classification task, achieving speeds up to 75x faster than previous state-of-the-art methods. Similarly, Lisi Chen et al. proposed Subtrajectory Alignment, which "stitches together" segments from different users to answer queries in sparse datasets where no single continuous record of a route exists.
By combining these perspectives—semantic infilling, physical constraints, and probabilistic modeling—researchers are transforming sparse, "holey" data into a coherent and actionable map of human and vehicle mobility.