MDPM: Modulating domain-specific prompt memory for multi-domain traffic flow prediction with transformers

Zhuang Zhuang, Lingbo Liu, Kan Guo, Xingtong Yu, Heng Qi, Yanming Shen, Baocai Yin

Abstract: Highlights•MDPM introduces a cross-domain spatial–temporal Transformer with an intra-modal ST2R encoder and inter-modal contrastive learning, enabling unified pattern modeling while preserving domain-specific features.•We devise distinct prompt-based spatial and temporal attention modules with node-level prompt vectors tailored to road network vertices and time steps for each dataset and layer.•We performed extensive experiments on six real-world datasets for both multi-step and single-step traffic prediction demonstrate superior performance over existing state-of-the-art (SOTA) methods.