STS2ANet: Spatio-Temporal Synchronized Sliding Attention Network for Accurate Cross-Day Origin-Destination Prediction

Haoli Wang, Jiangnan Xia, Yu Yang, Senzhang Wang, Jiannong Cao

Abstract: Accurately predicting Origin-Destination(OD) traffic flow is crucial in traffic planning, vehicle dispatching, user travel, etc. However, existing works mainly focus on modeling prolonged spatial-temporal trends of traffic flow, neglecting the divergence of spatial and temporal patterns at cross-day periods, especially the shift between weekdays and weekends. In this paper, we propose a Spatio-temporal Synchronized Sliding Attention Network (STS2ANet) to tackle this issue for accurate OD prediction. Specifically, we devise a Sliding Attention layer (SA) to learn the divergence of temporal traffic flow patterns at cross-day periods. Additionally, a Dynamic Graph Embedding module (DE) is proposed to properly learn the cross-day changes in spatial patterns of traffic flow. Notably, STS2ANet simultaneously learns the tightly coupled spatial-temporal patterns and their divergence over time, resulting in accurate OD prediction. Extensive experiments have been conducted in a real-world dataset, and the results demonstrate the performance superiority of STS2ANet against baselines.