Subgraph-aware graph structure revision for spatial-temporal graph modeling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Spatial-temporal graph modeling is crucial for fields like traffic and energy analysis.
• Existing methods struggle with feature smoothing and rely on potentially inaccurate predefined graph structures.
• The need for dynamic graph structures that capture true spatial correlations is evident.

Purpose of the Study:

• To propose a novel Subgraph-Aware Graph Structure Revision network (SAGSR) to address limitations in spatial-temporal graph modeling.
• To automatically infer and adapt dynamic spatial correlations, overcoming the constraints of static or manually constructed graphs.
• To prevent the obfuscation of spatial-temporal patterns by effectively handling homophilic and heterophilic node relationships.

Main Methods:

• Developed a subgraph-aware structure revision graph convolution module (SASR-GCM) to dynamically revise graph structures.
• Implemented a mechanism to separate graphs into homophilic and heterophilic subgraphs, using positive and negative aggregation weights respectively.
• Integrated a gated multi-scale temporal convolution module (GMS-TCM) for robust temporal modeling.

Main Results:

• The SAGSR network effectively captures complex spatial-temporal correlations.
• Experiments on traffic flow and energy consumption forecasting demonstrated superior performance compared to baseline methods.
• The approach successfully avoids pattern obfuscation by managing node feature aggregation.

Conclusions:

• The proposed SAGSR network offers a significant advancement in spatial-temporal graph modeling.
• Dynamic graph structure revision and subgraph-aware convolution are key to improving model accuracy.
• SAGSR shows strong potential for applications requiring accurate spatial-temporal predictions.