MTrajRec: Map-Constrained Trajectory Recovery via Seq2Seq Multi‑Task Learning

With the widespread use of GPS devices, massive low‑sampling‑rate trajectory data hinder the accuracy of navigation, travel‑time estimation, and driver‑behavior analysis. Traditional two‑step methods first up‑sample trajectories in free space and then match them to the road network, which introduces noise and is inefficient.

The authors propose MTrajRec, a Seq2Seq multi‑task learning model that jointly performs trajectory up‑sampling and map matching. The model predicts both the road‑segment ID and the proportion of movement on that segment, ensuring that generated points lie on the road network.

Problem Definition : Two trajectory formats are considered—(1) free‑space trajectories with latitude, longitude, and timestamp, and (2) map‑matched trajectories with road‑segment ID, proportion on the segment, and timestamp. The goal is to convert a low‑sampling‑rate free‑space trajectory into a high‑sampling‑rate map‑matched trajectory.

Model Structure : MTrajRec builds on a Seq2Seq encoder‑decoder architecture. The encoder uses a GRU to encode a grid‑based representation of the low‑sampling trajectory. The decoder employs multi‑task learning to first predict the road‑segment ID and then the movement proportion on that segment. Three enhancements improve performance:

Constraint mask: a distance‑weighted mask that limits predictions to road segments within 50 m, reducing information loss from grid quantization.

Attention mechanism: captures complex external factors such as weather, holidays, and POI information.

Feature module: integrates external contextual features with encoder outputs before decoding.

Experiments : Using taxi data from Jinan, the authors compare MTrajRec with three baselines on Precision, Recall, MAE, and RMSE. Ablation studies show that each of the constraint mask, attention, and feature modules contributes to performance gains. Visualizations demonstrate that MTrajRec’s predictions (blue points) align closely with ground truth (red points) and outperform baselines.

Conclusion : MTrajRec achieves end‑to‑end, map‑constrained trajectory recovery with higher accuracy and efficiency, validated on real‑world datasets. The combination of Seq2Seq multi‑task learning, constraint masking, attention, and contextual features addresses the main challenges of road‑network limitation, coarse grid representation, and external factor influence.