Existing research has focused on trajectory similarity computation under centralized data storage. However, in the real world, trajectories are often collected and held by multiple distinct entities that form a data federation. Due to commercial and legal restrictions, direct data sharing among these entities may be prohibited, which renders existing centralized methods incapable of supporting similarity computation over trajectory data federation. To address this challenge, a research team led by Professor Yongxin Tong published their new research on 15 May 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature. The team proposed a trajectory similarity computation framework based on federated learning, aiming to enable trajectory similarity analysis across data owners while ensuring the privacy and security of each party's data. Experiments on real datasets show that the proposed framework outperforms baselines in similar trajectory search and ��NN query.
The framework first coordinates multiple trajectory data owners in jointly training a federated embedding generation model based on federated learning. To tackle the challenge of data heterogeneity, a federated negative sample sharing mechanism is implemented to enhance the model's generalization capability. Subsequently, trajectory data are encoded into embedding vectors using this model. Finally, trajectory similarity over data federation is determined by computing the distances between these embedding vectors. This entire process ensures the effectiveness of trajectory similarity computation while preserving data privacy.
DOI:10.1007/s11704-025-41349-6