Spatial Structure-Aware Road Network Embedding via Graph Contrastive Learning

TL;DR: Experimental results on three downstream tasks over real-world road networks show that SARN outperforms state-of-the-art self-supervised models consistently and achieves comparable (or even better) performance to supervised models.

Abstract: Road networks are widely used as a fundamental structure in urban transportation studies. In recent years, with more research leveraging deep learning to solve conventional transportation problems, how to obtain robust road network representations (i.e., embeddings) applicable for a wide range of applications became a fundamental need. Existing studies mainly adopt graph embedding methods. Such methods, however, foremost learn the topological correlations of road networks but ignore the spatial structure (i.e., spatial correlations) which are also important in applications such as querying similar trajectories. Besides, most studies learn task-specic embeddings in a supervised manner such that the embeddings are sub-optimal when being used for new tasks. It is inecient to store or learn dedicated embeddings for every dierent task in a large transportation system. To tackle these issues, we propose a model named SARN to learn generic and task-agnostic road network embeddings based on self-supervised contrastive learning. We present (i) a spatial similarity matrix to help learn the spatial correlations of the roads, (ii) a sampling strategy based on the spatial structure of a road network to form self-supervised training samples, and (iii) a two-level loss function to guide SARN to learn embeddings based on both local and global contrasts of similar and dissimilar road segments. Experimental results on three downstream tasks over real-world road networks show that SARN outperforms state-of-the-art self-supervised models consistently and achieves comparable (or even better) performance to supervised models.