MLGO: Multi-Layer graph neural ODEs for traffic forecasting.

Abstract

Spatial-temporal graph neural networks have become increasingly prominent in traffic forecasting, driven by the inherently graph-structured nature of traffic networks. While temporal dependencies have been extensively studied, spatial correlations remain underutilized. Existing methods typically use predefined matrices that encode fixed spatial relations (e.g., spatial distances or traffic states similarities), or adaptive ones that capture task-specific latent patterns. However, relying on a single type of graph structure limits the ability to comprehensively capture the diverse spatial dependencies in traffic networks. To address this issue, we propose a general and extensible framework named Multi-Layer Graph neural Ordinary differential equations (MLGO), which integrates multiple complementary graph structures to enhance spatial representation. Specifically, MLGO employs a multi-layer graph architecture that incorporates a time-varying graph, a predefined road network, and an adaptively learned graph into a unified representation, where different graphs may compensate for or constrain one another, enabling richer and more expressive modeling of spatial correlations. Neural ordinary differential equations are further used to enable both inter-layer and intra-layer spatial aggregation, while ensuring continuity in temporal dynamics. Experiments on five real-world traffic datasets demonstrate that MLGO outperforms most state-of-the-art baselines and offers improved interpretability through the use of explicit and complementary graph structures.