DDGformer: Direction- and distance-aware graph transformer for traffic flow prediction

Abstract

Being an indispensable core technology of intelligent transportation systems, accurate traffic flow prediction contributes to improving people’s travel efficiency and safety. How to accurately model the dynamic spatio-temporal correlation in traffic data is a key challenge in traffic flow prediction. The models based on self-attention and Graph Neural Networks (GNN) have shown great potential in addressing this challenge. However, existing methods ignore the relative distance and direction of traffic flow series, and cannot accurately capture the real dynamic spatio-temporal correlations in traffic systems. To overcome this limitation, we propose a novel framework called Direction- and Distance-aware Graph Transformer (DDGformer). Specifically, it utilizes a self-attention module that simultaneously perceives direction and distance, enabling it to identify the relative position and direction of the original traffic flow series, and models long-term temporal correlations. Additionally, a dynamically enhanced adaptive graph convolution network is designed to capture dynamic traffic patterns in traffic systems. Extensive experimental results on four real-world traffic datasets show that our approach is overall more competitive and exhibits outstanding computational efficiency.