Semantic trajectory representation and retrieval via hierarchical embedding

Trajectory mining has gained growing attention due to its emerging applications, such as location-based services, urban computing, and movement behavior analyses. One critical and fundamental mining task is to retrieve specific locations or trajectories that satisfy particular patterns. However, existing approaches mainly represent the trajectory as a collection of geographic and temporal features, so the latent semantic properties are barely considered. In this paper, we introduce a new semantic trajectory representation method, which considers trajectory structures, temporal information, and domain knowledge to make efficient semantic retrieval possible. Specifically, we first introduce a synchronization-based model to identify multi-resolution regions of interest (ROIs) to extract structures from disordered raw trajectories. Afterward, we proposed a hierarchical embedding model to embed ROIs as well as trajectories on the hierarchical ROI network as continuous vectors by considering multiple kinds of semantic similarity. As a result, users can easily retrieve desirable ROIs or trajectories by computing the similarity among embedded vectors. Experiments show that our approach excels both classical trajectory metric-based models and state-of-the-art deep network embedding models in terms of retrieving interpretable ROIs and trajectories.