Trajectory similarity calculation, a cornerstone of trajectory data mining, is pivotal for diverse applications such as clustering, classification, and retrieval. While existing representation learning-based methods offer notable advantages in efficiency and accuracy, preserving the fidelity of similarity computation when processing large-scale trajectory data remains a significant challenge. To address this, this paper introduces a novel hybrid network architecture integrating Long Short-Term Memory (LSTM) and attention mechanisms to learn discriminative latent representations of trajectories. Moreover, we propose an Adaptive Contrastive Trajectory Learning (ACTL) module that dynamically refines the learning process through batch-adaptive temperature scaling and strategic hard negative mining, substantially improving boundary discrimination and robustness to data perturbations. Experimental validation on two real-world datasets, Porto and Chengdu, demonstrates the superiority of our model over state-of-the-art (SOTA) baselines in both similarity trajectory search and k-Nearest Neighbor (k-NN) query evaluations. The model exhibits exceptional performance, particularly under conditions of high noise and with large trajectory volumes, underscoring its practical applicability in demanding scenarios.
Ding et al. (Thu,) studied this question.