Beyond homophily in spatial-temporal traffic flow forecasting.

Area of Science:

• Artificial Intelligence
• Transportation Science
• Data Science

Background:

• Traffic flow forecasting is vital but complex due to spatial-temporal correlations.
• Existing Graph Convolutional Network (GCN) methods primarily use homophily graphs, limiting their ability to capture diverse road interactions.
• The homophily-heterophily dynamics of road networks, encompassing both close and distant correlations, have been largely overlooked.

Purpose of the Study:

• To propose a novel Homophily-Heterophily Spatial-Temporal Graph Convolution Network (H²STGCN) model.
• To address the limitations of existing methods by incorporating both homophily and heterophily in traffic flow forecasting.
• To capture diverse and dynamic node-wise correlations in road networks.

Main Methods:

• Utilized time-related node attributes to disentangle dynamic node-wise relations and derive homophily and heterophily Spatial-Temporal Graphs (STGs).
• Developed dual information propagation branches, each using a specific STG type with dilated causal spatial-temporal graph convolution operations.
• Introduced a Graph Collaborative Learning Module (GCLM) for mutual information transfer between branches.

Main Results:

• The proposed H²STGCN model effectively captures both homophilous and heterophilous spatial-temporal correlations.
• Dual information propagation branches exploit multi-range correlations from distinct perspectives.
• The Graph Collaborative Learning Module enhances information complementarity.

Conclusions:

• The H²STGCN model significantly outperforms state-of-the-art methods in traffic flow forecasting.
• Accounting for homophily-heterophily dynamics is crucial for accurate traffic modeling.
• The proposed approach offers a more comprehensive understanding of road interactions for improved forecasting.