TrafHILLM: Highway network traffic flow prediction with heterogeneous graph-based and instruction fine-tuned large language model

This paper presents TrafHILLM, a novel framework that integrates graph neural networks with large language model (LLM) for highway network traffic flow prediction. TrafHILLM constructs heterogeneous highway graphs incorporating both topological and functional relationships, then processes them through three aspects: 1) a unified architecture featuring adaptive spatiotemporal encoding via multi-relational graph convolutions and dilated temporal filters, 2) knowledge distillation through a unified sentence vector interface, and 3) two-stage LLM adaptation with traffic-specific instruction fine-tuning. Extensive experiments on three real-world benchmarks demonstrate TrafHILLM’s superior performance compared to state-of-the-art baselines. The integration of graph-based inductive biases enhances LLM’s adaptation to traffic data, leading TrafHILLM to have more stable training and improved prediction accuracy compared to pure sequence-based approaches. In addition, TrafHILLM not only improves prediction accuracy but also enhances interpretability and adaptability to complex highway network scenarios. Furthermore, the end-to-end training framework of TrafHILLM tailored for highway network traffic flow data, overcoming limitations related to data heterogeneity and high-dimensional feature spaces.