Traffic flow prediction via dynamic hypergraph learning.

Area of Science:

• Intelligent Transportation Systems (ITS)
• Machine Learning for Traffic Prediction
• Graph Neural Networks (GNNs)

Background:

• Accurate traffic flow prediction is crucial for intelligent transportation systems.
• Existing graph neural networks often overlook high-order relationships in traffic data.
• There is a need for models that capture complex spatial and temporal traffic dynamics.

Purpose of the Study:

• To propose a novel Transformer-based Hypergraph Convolutional Network (TSHGCN) for enhanced traffic flow prediction.
• To address the limitations of current methods in capturing high-order spatial correlations and global temporal features.
• To improve the accuracy and efficiency of predicting traffic flow evolution trends.

Main Methods:

• Utilized a hypergraph structure to model high-order nonlinear spatial correlations between traffic nodes.
• Developed an improved Transformer network incorporating time distillation and self-attention for global temporal feature extraction.
• Integrated spatiotemporal modeling using channel attention and multi-scale temporal information fusion for refined traffic flow representation.

Main Results:

• The TSHGCN model demonstrated superior performance on California datasets (PeMSD4 and PeMSD8) compared to state-of-the-art baselines.
• Achieved best results in core metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE).
• Statistical tests confirmed the significant performance improvement of TSHGCN under a unified experimental setting.

Conclusions:

• The proposed TSHGCN effectively captures high-order spatial and global temporal dependencies in traffic flow data.
• TSHGCN offers a significant advancement in traffic flow prediction accuracy and reliability for intelligent transportation systems.
• The model's ability to extract refined spatiotemporal features leads to statistically significant performance gains.