Effective Traffic Prediction with Self-Supervised Contrastive Learning

Abstract

Taxi demand prediction has recently attracted increasing research interest due to the growing availability of large-scale traffic data, which could empower various real-world applications. Accurate taxi demand prediction can improve vehicle utilization, reduce the time for passengers to wait for taxis, and mitigate traffic congestion. Although both spatial dependencies and temporal dynamics have been considered, most of the previous methods with over-complicated models might easily achieve suboptimal performance due to the overfitting issue. Contrastive unsupervised learning has recently shown encouraging progress, which has great potential to learn effective data representations without extensive manual labeling. In this paper, we utilize contrastive learning to construct an effective auxiliary task to learn feature representations of data in a self-supervised manner. The model learned via contrastive learning can be subsequently applied for downstream tasks, which is proven to be more robust against overfitting. The extensive experiments and evaluations on a large-scale dataset well demonstrate the superiority of our proposed model over other compared methods for taxi demand prediction.