Conditional Generative Adversarial Network-Based roadway crash risk prediction considering heterogeneity with dynamic...

Area of Science:

• Traffic Safety
• Machine Learning
• Data Science

Background:

• Roadway crash data are rare and random, posing challenges for real-time traffic safety management.
• Machine learning techniques are increasingly used for crash data augmentation to address limited sample sizes.
• Spatial and temporal variations necessitate incorporating specific crash data characteristics into augmentation and risk assessment.

Purpose of the Study:

• To develop a real-time crash risk model that accounts for heterogeneous crash data characteristics.
• To improve the accuracy of crash prediction models by addressing data imbalance issues.
• To identify key variables influencing crash risk in different spatial and temporal contexts.

Main Methods:

• Crash data were clustered to identify distinct risk situations.
• Boruta-SHAP (explainable artificial intelligence) identified key predictive variables.
• Conditional Generative Adversarial Network (CGAN) augmented clustered crash data, preserving cluster-specific characteristics.
• Developed and compared various crash risk models, including XGBoost, BLM, RF, and SVM.

Main Results:

• The CGAN-based XGBoost model demonstrated superior performance compared to other models.
• Temporal speed difference (10-minute intervals) and precipitation were identified as significant predictors of crash risk.
• The study successfully addressed data imbalance issues in crash and non-crash datasets.

Conclusions:

• Distinguishing crash risk characteristics is crucial for accurate crash prediction.
• The proposed method effectively handles data imbalance in traffic safety analysis.
• This approach offers valuable insights for real-time traffic safety management and predictive modeling.