Dynamic graph neural network with adaptive edge attributes for air quality prediction: A case study in China

Area of Science:

• Environmental Science
• Computer Science
• Data Science

Background:

• Air quality prediction is a complex spatiotemporal modeling challenge.
• Existing methods often handle spatial and temporal dependencies in isolation.
• Recurrent Neural Networks (RNNs) neglect spatial information, while Graph Convolutional Networks (GCNs) require pre-defined spatial structures.

Purpose of the Study:

• To develop a novel model for accurate air quality prediction.
• To overcome the limitations of existing methods that rely on prior information for spatial relationships.
• To dynamically learn spatial dependencies without human pre-defined structures.

Main Methods:

• Proposed a Dynamic Graph Neural Network with Adaptive Edge Attributes (DGN-AEA).
• Employed a message passing network to generate an adaptive bidirected dynamic graph.
• Learned edge attributes as model parameters through end-to-end training, eliminating the need for prior information.

Main Results:

• The DGN-AEA model achieved state-of-the-art performance compared to baseline models.
• The model successfully learned adaptive edge information without prior structural input.
• Identified hidden structural information between stations as a valuable byproduct.

Conclusions:

• The DGN-AEA model offers a more effective approach to spatiotemporal air quality prediction.
• Dynamic graph learning eliminates the need for manual feature engineering of spatial correlations.
• The model's ability to uncover hidden structures aids in decision-making analyses.