LATTE: Learning Adaptive Segmentation for Efficient and Effective Trajectory Similarity Learning

Trajectory similarity learning (TSL) models represent variable-length trajectories as fixed-size vectors, for which the similarity is much easier to compute. Recent TSL models often map-match raw GPS trajectories onto road networks to exploit richer information. Although effective, such models rely on heavy modeling structures for both road networks and trajectories, leading to overfitting and inefficiency, especially on long trajectories with many trajectory segments. In this work, we propose LATTE, a novel framework that Learns Adaptive segmenTaTion for Efficient and effective TSL. LATTE uses a lightweight road segment embedding module that first decouples topological modeling from road segment feature modeling during training and then decouples the whole road segment embedding from inference. Also, LATTE includes a learned adaptive segmentation module to merge consecutive trajectory segments into sub-trajectories and substantially reduce the trajectory length. To train LATTE efficiently and effectively, we design a tailored knowledge distillation strategy. It first trains a teacher model without trajectory segmentation and then trains the complete LATTE model on adaptively segmented trajectories to mimic the teacher, thereby inheriting strong representation capability with lower computational cost. We compare LATTE with six state-of-the-art baselines on two real datasets. The results show that LATTE outperforms the best baselines by up to 34.6% in TSL effectiveness with an up to 11.9× speedup in inference efficiency.