RGDAN: A random graph diffusion attention network for traffic prediction.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Network Science

Background:

• Traffic prediction relies on graph structures, but road networks are complex with variable temporal features.
• Current methods use fixed weights (e.g., distance) and overlook road characteristics and traffic flow correlations.
• Existing models often ignore global spatial dependencies and struggle with information extraction at limited graph depths.

Purpose of the Study:

• To develop an advanced model for accurate traffic prediction.
• To address limitations in spatial feature extraction and temporal dependency modeling.
• To improve the precision of traffic flow forecasts in complex road networks.

Main Methods:

• Proposed a novel Random Graph Diffusion Attention Network (RGDAN).
• RGDAN integrates a graph diffusion attention module for adaptive spatial weight learning.
• Incorporated a temporal attention module to capture temporal correlations.

Main Results:

• RGDAN demonstrated superior performance on three large-scale public datasets.
• Achieved 2%-5% higher prediction precision compared to state-of-the-art methods.
• Effectively captured both local and global spatial dependencies and temporal correlations.

Conclusions:

• RGDAN offers a significant advancement in traffic prediction accuracy.
• The model's adaptive weighting and attention mechanisms enhance spatial and temporal feature extraction.
• RGDAN provides a more robust solution for complex transportation network analysis.