Introduction
Generating realistic synthetic human trajectories is a critical task for urban planning, traffic management, and privacy-preserving data analysis. However, human movement is notoriously difficult to model accurately. It is high-dimensional and deeply personal, yet it follows underlying semantic patterns, such as the daily commute. Traditional models often struggle to balance these local geographic constraints with global semantic logic.
Researchers from the National University of Singapore and SAP have proposed a shift in perspective: moving away from sequential "step-by-step" models toward a holistic, non-parametric approach that treats trajectories as spatial structures.
The Problem with Sequential Modeling
Most existing mobility models are sequential (or "time-major"). They predict the next location based on a short history of previous visits using techniques like Markov Chains or LSTMs. While effective for short-term predictions, these models face two major challenges:
- Exposure Bias: Errors compound over long trajectories, leading to significant declines in accuracy as the sequence progresses.
- Semantic Blindness: Sequential models often fail to capture the global context of a trip—the "why" behind the movement. For instance, they may struggle to correlate a morning departure from home with an evening return, focusing instead on immediate geographic transitions.
The challenge is to generate a whole trajectory in a consistent manner that respects both the physical constraints of geography and the logical constraints of human behavior.
Methodology: Trajectories as Images
The core innovation of this work is the location-major representation. Instead of modeling a sequence of coordinates over time, the model represents a trajectory by asking when and for how long a person visited specific locations.
The Generative Pipeline
- Grid Discretization: The geographic area is divided into an $N_1 \times N_2$ grid.
- Trajectory Embedding: Each trajectory is transformed into a 3D tensor ($N_1 \times N_2 \times K$). Each cell $(x, y)$ stores information about the $k$-th visit to that location, specifically the start time ($t$) and duration ($d$). This turns a path into a sparse "map" or "spatial drawing."
- GAN Architecture: The model employs a Wasserstein GAN with Gradient Penalty (WGAN-GP). By using convolutional layers, the GAN naturally detects local spatial patterns, recognizing that visits to adjacent grid cells are highly correlated.
- Transformation: Synthetic tensors generated by the GAN are post-processed to convert the spatial representation back into standard sequential trajectory data.
Experimental Results
The model was evaluated using the Nokia Lausanne location trajectories dataset, comprising approximately 2.6 million trajectories. It was compared against several baselines, including First-order Markov Chains (MC), Hidden Markov Models (HMM), and LSTMs trained with Maximum Likelihood Estimation (LSTM-MLE).
Quantitative Fidelity
The researchers used Jensen-Shannon Divergence (JSD) to measure the similarity between synthetic and real data distributions for visiting probability $p(r)$, visit time $p(r,t)$, and duration $p(r,d)$.
The proposed method achieved significantly lower JSD scores than all baselines. For example, in the visiting probability distribution $p(r)$, the model achieved a JSD of 0.066, vastly outperforming the LSTM-MLE baseline, which scored 0.386.
Semantic Preservation
Beyond raw geographic accuracy, the model demonstrated superior "semantic" performance. It was more precise at identifying the top-N most popular locations and maintained the "semantic closeness" (functional logic of movement) found in real-world data, as evidenced by pairwise semantic distance distributions. This suggests the model captures the underlying social and functional logic of human movement better than traditional transition-based models.
Conclusion and Future Work
By treating trajectories as static spatial objects rather than sequences, this non-parametric approach bypasses the compounding errors of traditional models and captures high-order mobility motifs.
However, the model has limitations, including sensitivity to grid resolution and a fixed maximum number of visits per cell ($K$). Future research directions include addressing these discretization constraints and incorporating Differential Privacy to ensure that synthetic data generation does not compromise the privacy of individuals in the training set.
Summary updated