Spatio-Temporal Balanced Hypergraph Network for cross-city urban traffic flow prediction under data scarcity

Urban traffic flow prediction is crucial for intelligent transportation systems, but existing spatio-temporal forecasting methods often degrade in cross-city scenarios due to limited target-city data. Although cross-city transfer learning provides an effective solution, two challenges remain: insufficient representation of high-order regional knowledge and source-domain distribution imbalance that biases learning toward frequent patterns. To address these issues, we propose a knowledge-supported Spatio-Temporal Balanced Hypergraph Network (STBH) under a source-city-driven pretraining paradigm, which explicitly represents and reuses urban functional, dynamic relational, and source-domain distribution knowledge. Specifically, the Region Traffic Pattern Extractor (RTPE) and Multi-scale Dynamic Hypergraph Spatio-Temporal Predictor (MDHSTP) capture transferable regional structures and time-varying interactions through static semantic and multi-scale dynamic hypergraphs. The Region-Group Density-aware Balanced Learning mechanism (RGBL) estimates individual- and group-level densities and converts distribution knowledge into adaptive weights to reduce biased transfer. The Cross-space Alignment and Bidirectional Optimization mechanism (CABO) couples static and dynamic knowledge via structure-aware alignment and statistical consistency constraints. Experiments show that STBH outperforms state-of-the-art methods in both few-shot and zero-shot cross-city scenarios, improving prediction accuracy and generalization while providing a scalable knowledge representation and reuse framework for traffic engineering decision support.