Trajectory Similarity Measurement With Spatial-Temporal Graph Contrastive Learning for Traffic Networks

Trajectory similarity computing enables us to gain deeper insights into the movement patterns of objects, which benefits intelligent transportation system applications from urban planning to travel recommendation. In this study, we explore a new Spatial-Temporal Graph Contrastive Learning method for Trajectory Similarity computation (STGCL-TrajSim). We design two contrastive learning methods on the road network trajectory. First, time contrastive learning employs two data augmentation methods and efficient sampling strategies to capture multi-scale temporal information. Secondly, spatial contrastive learning adopts an importance-based graph augmentation strategy to preserve both the graph topology and spatial structure information. Thirdly, we use the road transition probability matrix to represent the road access frequency, and adopt a time interval matrix to reflect traffic congestion conditions, and design a co-attention mechanism to capture the interactions between temporal and spatial features. Finally, experimental results on real datasets show that our model outperforms state-of-the-art competitors consistently and achieves better performance in terms of accuracy.