Solution to data imbalance and complex interactions in traffic conflict modeling: a hypergraph and generative AI...

Area of Science:

• Traffic Safety Engineering
• Machine Learning Applications
• Intelligent Transportation Systems

Background:

• Current traffic conflict models struggle with imbalanced data and dynamic interactions.
• Limitations in existing models affect generalization and real-world applicability.
• Need for robust methods to predict diverse traffic conflict patterns.

Purpose of the Study:

• To develop an enhanced traffic conflict prediction model.
• To improve handling of imbalanced datasets in traffic conflict analysis.
• To identify key features influencing traffic conflicts.

Main Methods:

• Utilized an enhanced Two-dimensional Time-to-collision (2D-TTC) metric with vehicle interaction relationships.
• Employed undersampling and oversampling techniques, including a generative adversarial network with self-attention.
• Compared various machine learning and deep learning models, focusing on the hypergraph attention network (HGAT) with Shapley additive explanations (S-HGAT).

Main Results:

• The enhanced model achieved an F1-score of 94.21%, significantly improving from 76.35% with undersampling alone.
• The S-HGAT model demonstrated superior learning capability.
• Vehicle speed was identified as the most influential factor; a top six feature set yielded an F1-score of 98.41% and 97.66% accuracy.

Conclusions:

• The proposed method effectively addresses data imbalance and enhances traffic conflict prediction accuracy.
• The S-HGAT model and identified key features offer a robust approach for intelligent transportation systems.
• Findings have significant implications for improving road safety and traffic management.