CRRL: Contrastive Region Relevance Learning Framework for Cross-city Traffic Prediction

Abstract

Accurate traffic prediction plays a pivotal role in smart urban systems by enabling effective traffic management and improving the quality of life for residents. While intra-city traffic prediction has been extensively studied, cross-city traffic prediction remains a challenging task due to data scarcity in target cities, domain gaps between cities, and the risk of negative transfer. To address these challenges, we propose a novel Contrastive Region Relevance Learning (CRRL) framework. CRRL leverages contrastive learning to align region-level spatiotemporal patterns and transfer high-quality knowledge between cities. Specifically, CRRL integrates three key modules: (1) a Dual-branch Spatiotemporal Encoder (DSE) to capture region-pair and high-order region group relationships, (2) a Pseudo-Label Generation (PLG) module for aligning cross-city embedding similarities, and (3) a Reliable Region Selection (RRS) module for contrastive learning within consistent regions. Extensive experiments on real-world datasets demonstrate that CRRL achieves state-of-the-art performance in cross-city traffic prediction under data-scarce scenarios, showcasing its practicality and effectiveness in addressing urban traffic challenges.