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

Abstract

Multi-domain traffic flow prediction aims to develop a versatile model that uses historical traffic data from various sources to forecast future traffic conditions across these individual datasets. Existing deep traffic prediction models typically focus on mining spatial–temporal relationships in a single dataset. However, there are two limitations should be considered: Lack of Model Universality, current traffic prediction research remains constrained by the absence of a universal model adaptable to multiple datasets, restricting performance improvement across diverse scenarios; Underutilized Cross-Dataset Similarities, while existing datasets exhibit both exclusive and shared spatial–temporal patterns, effectively leveraging these common patterns to enhance model performance continues to present technical challenges. To overcome the limitations mentioned above, this study introduces a straightforward yet efficient Modulating Domain-Specific Prompt Memory (MDPM) to model complex spatial–temporal interaction and better leverage similar spatial–temporal patterns across diverse datasets. Specifically, our approach is tailored with three key innovations: (1) A domain-shared encoder incorporating intra-modality Spatial–Temporal Rotary Position Encoder (ST2R) to capture universal patterns; (2) A gate fusion mechanism enhanced by contrastive learning with inter-modality ST2R to optimize spatial–temporal feature alignment; (3) Domain-specific learnable prompt vectors that dynamically guide each transformer layer in capturing unique urban traffic characteristics at node-level temporal granularity. Notably, this architecture achieves state-of-the-art performance without requiring supplementary road network data. Comprehensive experiments conducted on six real-world public traffic datasets show that our proposed method significantly surpasses existing state-of-the-art approaches.