Traffic Conflict Prediction: S-HGAT Model Hits 98% Accuracy

98.41 per cent F1-score. That is the new benchmark for safety modelling. By combining hypergraph attention networks (HGAT) with Shapley additive explanations, researchers have effectively solved the issue of data imbalance in crash forecasting.

Solving the Imbalance in Traffic Conflict Prediction

Safety models often fail because crashes are rare. Algorithms trained on hours of safe driving struggle to recognise the split-second anomaly of a near-miss. This study tackled that asymmetry head-on. The team employed a generative adversarial network (GAN) with self-attention layers to synthesise data for minority class samples.

The results were immediate. Using undersampling alone yielded a mediocre F1-score of 76.35 per cent. Adding the GAN-generated data elevated this to 94.21 per cent. Synthetic data bridged the gap between theoretical modelling and real-world applicability.

The S-HGAT Advantage

Standard machine learning models view vehicles as isolated points. They are not. Traffic is a cohesive, interactive flow. The Hypergraph Attention Network (S-HGAT) captures these complex relationships better than deep learning alternatives. It maps the dynamic interactions between multiple vehicles simultaneously, offering a superior grasp of the road environment.

Key Factors Driving Risk

Speed dominates. It is the single most influential variable. However, accuracy requires context. A comprehensive re-evaluation identified the top six features necessary for high-fidelity traffic conflict prediction:

• Vehicle speed.
• Number of vehicles ahead.
• Standard deviation of speed within the flow.
• Average speed of the traffic stream.
• Distance to road markings.
• Peak traffic hour indicators.

Focusing on these specific inputs allowed the S-HGAT model to reach a final accuracy of 97.66 per cent. This precision is essential for the deployment of autonomous systems and advanced driver-assistance technologies.