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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Traffic Engineering

Background:

• Spatial-temporal graph neural networks are key for traffic forecasting due to network structure.
• Existing models underutilize spatial correlations, often using single, fixed, or adaptive graph structures.
• This limitation hinders comprehensive modeling of diverse spatial dependencies in traffic networks.

Purpose of the Study:

• To propose a novel framework, Multi-Layer Graph neural Ordinary differential equations (MLGO), for enhanced spatial representation in traffic forecasting.
• To integrate multiple complementary graph structures for a more robust spatial modeling approach.
• To improve the accuracy and interpretability of traffic forecasting models.

Main Methods:

• Developed a multi-layer graph architecture integrating time-varying, predefined road network, and adaptive graphs.
• Employed neural ordinary differential equations for inter-layer and intra-layer spatial aggregation.
• Ensured temporal continuity within the spatial modeling framework.

Main Results:

• MLGO demonstrated superior performance compared to state-of-the-art baselines across five real-world traffic datasets.
• The integrated multi-layer graph approach effectively captures complex spatial correlations.
• The framework offers improved interpretability through explicit and complementary graph structures.

Conclusions:

• MLGO provides a general and extensible solution for leveraging diverse spatial information in traffic forecasting.
• The integration of multiple graph structures significantly enhances spatial-temporal modeling capabilities.
• The proposed method represents a significant advancement in accurate and interpretable traffic prediction.