Abstract The core challenge of spatio-temporal traffic flow prediction lies in accurately modeling the bidirectional coupling relationship where temporal trends modulate spatial correlations and spatial events feedback to temporal evolution. Existing models fail to capture these intrinsic dynamics because they rely on unidirectional dependency modeling or static fusion strategies. This paper proposes a Multi-Scale Bidirectional Coupling Modulation Transformer (BiCoMT), which enables end-to-end modeling of spatio-temporal coupling through coordinated module design. The core contributions are as follows: (1) A bidirectional coupling modulation mechanism is proposed to address the limitations of traditional unidirectional dependency modeling, achieving dynamic modulation of space by time and feedback guidance of time by space, constructing a closed-loop coupling modeling logic; (2) A collaborative architecture of Multi-Scale Temporal Convolutional Network (MSTCN) and Dynamic Topology Graph Convolutional Network (DTGCN) is designed, where the former extracts cross-scale temporal features and the latter generates time-modulated dynamic spatial dependencies, with both achieving bidirectional dynamic interaction of spatio-temporal features through the Sample-Node Adaptive Fusion (SNA Fusion) module; (3) A causal sparse attention encoder is introduced to efficiently model long-range bidirectional coupling dependencies while ensuring temporal causality, balancing accuracy and efficiency. Experiments on multiple public traffic datasets demonstrate that BiCoMT achieves competitive or superior performance compared to current SOTA models while substantially reducing parameter count and inference time, providing high-accuracy, low-latency technical support for dynamic scheduling in Intelligent Transportation Systems (ITS).
Yi et al. (Tue,) studied this question.