MATIBL: Multi-level Augmented Trajectory Imputation via Bootstrapping Learning

Trajectory imputation plays a vital role in intelligent transportation systems. However, real-world trajectory data often suffers from noise, missing segments, and limited labeled samples, which significantly hampers the performance of existing methods. Moreover, conventional approaches typically depend on large-scale labeled datasets or negative samples, limiting their scalability to large or sparsely annotated datasets. To address these challenges, we propose MATIBL, a Multi-level Augmented representation learning framework for Trajectory Imputation via Bootstrap Learning. MATIBL enhances trajectory representation learning through trajectory-specific data augmentations and task-aware objectives. Specifically, a dual-branch bootstrap pretraining strategy is employed to learn robust representations without requiring negative samples. A mixed augmentation scheme incorporating random dropout and Gaussian noise simulates real-world trajectory irregularities, while multi-faceted features-including spatial, temporal, and attribute information-enable comprehensive modeling. A BERT-based encoder supports position and order prediction tasks to capture fine-grained spatial-temporal dependencies. Extensive experiments on two large-scale real-world trajectory datasets (Chengdu and Porto) demonstrate that MATIBL consistently outperforms state-of-the-art methods. It maintains high accuracy under extreme missing rates (up to 90%) and exhibits strong generalization across diverse datasets, validating its effectiveness and practicality.