Predicting road traffic accident severity from imbalanced data using VAE attention and GCN

Area of Science:

• Road safety and traffic management
• Machine learning for predictive analytics
• Data science for social security

Background:

• Traffic accidents pose significant social security risks, necessitating accurate severity prediction.
• Imbalanced datasets, with few major accident samples, hinder effective prediction models.
• Existing methods struggle to capture the importance of minority samples in accident prediction.

Purpose of the Study:

• To develop an advanced traffic accident severity prediction method.
• To address the challenge of imbalanced sample data in accident prediction.
• To enhance the accuracy and reliability of predicting severe traffic accidents.

Main Methods:

• Utilized Variational Autoencoders (VAE) for generating minority accident samples.
• Integrated a self-attention mechanism with VAE to capture latent feature dependencies.
• Employed Graph Convolutional Networks (GCN) with a swish function for topological structure and nonlinear characteristic extraction.
• Applied smooth L1 loss function for improved model optimization with integer accident data.

Main Results:

• The proposed method significantly improves traffic accident severity prediction accuracy by 20%.
• Demonstrated high accuracy in predicting serious traffic accidents.
• Generated synthetic data that aligns with real-world accident characteristics.

Conclusions:

• The VAE and GCN-based method offers a robust solution for imbalanced traffic accident data.
• Accurate prediction of severe accidents can be achieved, aiding traffic safety management.
• The findings provide valuable insights for traffic safety systems and decision-makers.