Robust Spatial-Temporal Similar Trajectory Search via Structure-Enhanced Domain-Invariant Learning

Similar trajectory search is a fundamental task in spatial-temporal data analysis, supporting applications such as route planning, traffic forecasting, and mobility behavior understanding. However, existing methods often struggle to generalize across domains. In this paper, we address the problem of robust spatial-temporal similar trajectory search, which aims to generalize across multiple domains or cities, and is applicable to nonuniform trajectory data with irregular time intervals from diverse sources. This task is particularly challenging when deploying such systems across multiple cities or regions, where urban layouts, road networks, and mobility behaviors can differ dramatically. To tackle this problem, we propose a novel Robust similar Trajectory search (RoTraj) framework that introduces structure-enhanced domain-invariant learning to improve cross-domain generalization for non-uniform data. Specifically, RoTraj aligns variablelength trajectories through dynamic spatial-temporal dependency modeling, and employs an adaptive domain discriminator to learn domain-invariant representations for effective knowledge transfer. Moreover, it incorporates a structure-enhanced intradomain dynamics learning module that captures intrinsic structural similarities despite geometric discrepancies, leading to more discriminative and robust trajectory embeddings for similarity computation. Extensive experiments demonstrate that RoTraj significantly outperforms existing approaches with 72.6% improvement on average. Implementation details and code are publicly available at https://anonymous.4open.science/r/RoTraj5022.