Progressive Decomposition-Enhanced Time-Varying Graph Neural Network for Traffic Forecasting

Abstract

Traffic forecasting, a core technology in intelligent transportation systems, has a broad range of applications. The fundamental challenge in traffic prediction lies in effectively modeling the complex spatio-temporal dependencies inherent in traffic data. Spatio-temporal graph neural network (GNN) models have emerged as one of the most promising approaches to address this challenge. However, GNN-based models for traffic forecasting have two significant limitations: i) Most methods model spatial dependencies in a static manner (predefined or self-learning), which fails to capture the time-varying nature of spatial dependencies in real-world scenarios; ii) It is unreliable to capture temporal and spatial dependencies in entangled temporal patterns. To this end, we propose a P rogressive D ecomposition-enhanced T ime- V arying G raph N eural N etwork, namely PDTVGNN, for accurate traffic forecasting. Specifically, we design a time-varying graph generator that incrementally generates a series of adjacency matrices to capture the time-varying spatial relationships. Moreover, we adopt a novel progressive decomposition idea where the decomposition blocks are embedded as internal blocks to decouple the entangled temporal patterns gradually. The decoupled trend and seasonal parts are modeled via the proposed spatio-temporal normalization module and attention mechanism, respectively. Extensive experimental results on four real-world public traffic datasets demonstrate that the proposed method outperforms state-of-the-art baselines.