Multi-feature hybrid network for traffic flow prediction based on mobility patterns

Abstract

Future flow prediction in spatiotemporal traffic data is a critical requirement for real-world applications, particularly for multi-feature and large-scale data with intricate forecasting mechanisms and varied predictability. Prior sequence-to-sequence studies have demonstrated the superiority of attention learning in prediction tasks by effectively capturing reliable dependencies/correlations. However, despite the promising results, the single-feature input pattern of traffic prediction causes an information utilization bottleneck. In this paper, we focus on discovering the cause-effect relationship of traffic dynamics by fusing the origin–destination (OD) flow. To achieve both high accuracy and universality, we design a scalable composition architecture—a Multi-Feature Hybrid Network (MFHN)—based on the existing framework of spatiotemporal feature modeling. In particular, we break with the preprocessing convention of feature composition and propose a Hybrid-Correlation mechanism by integrating similar subseries of the OD flow. Furthermore, inspired by graph learning, we introduce the mobility pattern based on the OD flow, which reveals the node-to-node dependencies. In the experiment, we consider various prediction tasks with state-of-the-art baseline models and find that the MFHN yields competitive accuracy in short- and long-term prediction