Hypergraph mixer WaveNet: A lightweight spatio-temporal framework utilizing hypergraphs for high-order spatial modeling in traffic forecasting.

Abstract

Traffic speed forecasting plays a critical role in modern Intelligent Transportation Systems (ITS) by enabling accurate monitoring and prediction of traffic conditions for efficient traffic management. However, the inherent spatio-temporal complexity of traffic data poses significant challenges to the development of highly accurate forecasting models. Existing approaches typically model traffic data on ordinary graphs that reflect the physical topology of transportation networks. While effective, such graph-based methods are limited to capturing low-order (pairwise) relationships, which restricts their ability to exploit the rich high-order dependencies present in real-world traffic systems. To address this limitation, we adopt an edge-dependent vertex-weighted hypergraph representation and introduce a k-hop proximity neighbor modeling strategy to transform ordinary graphs into hypergraphs. Based on this representation, we propose a novel framework, termed the Hypergraph Mixer WaveNet (HGMWN), which efficiently captures high-order spatial dependencies with a compact architecture. HGMWN integrates a Dilated Temporal Convolutional Network within the WaveNet framework to model both short-term and long-range temporal dependencies without increasing network depth. Moreover, the architecture incorporates dedicated modules to jointly model static structural relationships and dynamic diffusion processes on the hypergraph, enabling more expressive representations of complex spatial interactions. The experiments conducted on three real-world traffic datasets demonstrate that HGMWN consistently outperforms recent state-of-the-art models. HGMWN achieves 4%-5% improvements in MAE while keeping a compact model size and low inference latency compared to existing spatio-temporal graph neural networks.