STGCNFormer: Spatio-Temporal Dual-Stream Graph Convolutional Networks and Transformers for Traffic Forecasting

In intelligent transportation systems, traffic prediction constitutes the fundamental component of traffic optimization, aiming to precisely forecast traffic flow. Modeling spatio-temporal interactions is a crucial step in achieving long-term and accurate predictions. Traditional methods are mostly based on single-stream networks, which alternately extract spatio-temporal features through sequential iterations. However, time and space are inherently two different modalities. Employing a single-stream network to address multi-modal tasks is suboptimal. In this paper, we propose Spatio-Temporal Dual-Stream Graph Convolutional Networks and Transformers (STGCNFormer), a dual-stream network for traffic forecasting, decoupling spatio-temporal features in parallel through a temporal stream and a spatial stream. The core of STGCNFormer is a STGCNFormer block, comprising three components. The first component is a GCN block, which restricts the kernel of graph convolution and adopts Chebyshev polynomials of the first kind to approximate the convolutional kernel, capturing spatial interactions and achieving localized convolution. Secondly, we present a Transformer block, designed as a simplified version of vanilla Transformers, which captures temporal interactions and significantly reduces computational complexity. Thirdly, we propose a cross attention mechanism, consisting of a temporal attention mechanism and a spatial attention mechanism, which serves as a bidirectional bridge for sharing information between the dual streams. Experimental results demonstrate that STGCNFormer achieves superior prediction accuracy with reduced computational complexity compared with previous state-of-the-art methods.