PSformer: Periodic-aware Semantic Transformer for Traffic Prediction

Traffic prediction plays an important role in Intelligent Transportation Systems (ITS). The main challenge lies in effectively capturing the dynamic multiple temporal periodic correlations and the long-range spatial correlation of traffic data. Despite the significant progress of many existing works, these methods often have two major limitations: 1) They mined the dynamic multi-period properties by using raw traffic sequences or the fixed periodicity strategy (e.g., hours, days, weeks), which failed to capture the dynamic multi-period characteristics of temporal correlation. 2) They mined the long-range spatial correlation of traffic data by stacking multilayer networks or directly using traditional similarity algorithms (e.g., conventional DTW). However, DTW has its own limitations leading to sub-optimal similarity assessment. To address these issues, we propose a periodic-aware spatial semantic transformer called PSformer for traffic prediction. Specifically, we propose the Periodic-aware Embedding Module (PAEmbed) to capture the dynamic multi-period properties by decoupling the traffic sequence into the multilevel frequency components via Fast Fourier Transform (FFT). In addition, we propose a Semantic Spatial Attention Mechanism (SSAM) to capture the long-range spatial correlation. In SSAM, we propose Time-weighted Dynamic Time Warping (TDTW) to model spatial correlations in semantically identical but geographically distant regions, which avoids considering two traffic patterns with large time spans as similar. Finally, to evaluate the performance of PSformer, we conduct extensive experiments on four real datasets. Experimental results show that our model achieves better performance than other state-of-the-art methods.