Cross-city traffic state prediction based on knowledge transfer framework

Abstract

Real-time and accurate traffic state prediction is crucial for smart transportation systems in cities, serving as the foundation for various downstream intelligent mobility applications. However, existing traffic prediction models heavily rely on large-scale traffic data, which require significant cost expenses. Therefore, we propose a novel knowledge transfer framework to achieve accurate prediction of cross-city traffic state and to reduce expenses. Firstly, a feature extraction module based on a shared processing mechanism is proposed to achieve the intrinsic connection of multi-graph networks, which uses multi-layers graph convolution and instance attention layer to extract embedding feature. Secondly, a feature matching module based on Linear Transformer is proposed to achieve feature adaptation between the source city and the target city, which utilizes self-attention and interaction attention mechanism to match and generalize the embedding feature. Then, a joint meta-learning module is used for the pre-training of traffic prediction model, which includes source training, target fine-tuning, and parameters updating. Finally, the final fine-tuning of the traffic prediction model is carried out on target city to achieve accurate prediction of cross-city traffic state. Experimental results conducted on real traffic datasets demonstrate that the model framework proposed in this paper outperforms baseline models in terms of performance.