A Katz-Attention Enhanced Decomposition-Fusion Framework for Robust Traffic Forecasting

Predicting traffic data is crucial for public safety and efficient resource allocation. Due to the constantly changing nature of traffic data over temporal and spatial patterns, traffic forecasting remains highly challenging. Various spatial-temporal network architectures are employed to effectively process and analyze interconnected spatial and temporal data patterns. However, existing work has mainly focused solely on learning the spatial-temporal dependencies in the data without addressing the issue of distribution shift in non stationary traffic data. Additionally, there is currently a lack of an efficient model to adequately extract spatial dependencies in the data. This study introduces a novel katz-attention enhanced decomposition-fusion approach (KADF) to mitigate the distribution shift challenge which can improve model generalization capability and enhance model robustness.Our method separates intricate traffic data into stable patterns and variable occurrences, then employs a two-pronged spatial temporal neural network to independently analyze these paterns. The reasonable future traffic data can be predicted based on fused trend and event. In addition, we propose a method that utilizes pre-defined graph structure information to model spatial dependencies. Compared to traditional graph attention networks, this method greatly reduces computational complexity. Our research involved comprehensive testing across four distinct datasets, which highlighted the advantages of our model. The results indicated improved forecast precision and reduced computational costs compared to existing approaches.