CLEAR: Ranked Multi-Positive Contrastive Representation Learning for Robust Trajectory Similarity Computation

Abstract

Similarity computation is the core building block for GPS trajectory analyses. Nevertheless, due to the inherent limitations of GPS technology and devices, similar trajectories may have noises and low sampling rates, resulting in being inaccurately considered dissimilar. To fortify the robustness of trajectory similarity computation, we propose a novel contrastive representation learning framework (CLEAR). We adaptively combine spatial information with sequential information to model essential properties of trajectory data. Subsequently, we rank multiple positive instances (i.e., different variations of an anchor trajectory) based on their similarities to the anchor instance. We propose a specialized loss function that strategically harnesses these positive instances, iteratively associating harder positive instances with higher rank values. Moreover, we propose a multiple augmentation strategy to generate and utilize multiple positive instances. We conduct extensive experiments on two real-world trajectory datasets. The results validate the superiority of CLEAR over state-of-the-art models in terms of robust trajectory similarity computation against noises and low sampling rates.