An efficient methodology for modeling imbalanced traffic crashes through deep learning techniques

Area of Science:

• Traffic Safety Engineering
• Machine Learning
• Data Science

Background:

• Accurate crash injury severity prediction is essential for traffic safety.
• Imbalanced crash datasets, where severe injuries are rare, pose challenges for predictive models.
• Traditional resampling techniques often have limitations in addressing class imbalance.

Purpose of the Study:

• To develop a hybrid methodology (ENN-CTGAN) to address class imbalance in crash injury severity prediction.
• To compare the performance of a hybrid LSTM-GRU model with other machine learning models using the proposed methodology.
• To identify optimal synthetic data ratios and evaluate data quality using a novel framework.

Main Methods:

• Developed a hybrid methodology named ENN-CTGAN for class imbalance.
• Implemented a hybrid LSTM-GRU model for injury severity prediction.
• Investigated synthetic data ratios (1:1, 1:2, 1:4, 1:6) and compared ENN-CTGAN with SMOTE, Random Oversampling, and Random Undersampling.

Main Results:

• The ENN-CTGAN framework outperformed conventional resampling methods across various predictive models.
• The hybrid LSTM-GRU model achieved the best performance (G-mean = 0.5452) within ENN-CTGAN at a 1:1 ratio.
• The XGBoost classifier achieved the highest overall performance (G-mean = 0.5643) with ENN-CTGAN at a 1:4 synthetic data ratio.

Conclusions:

• The ENN-CTGAN methodology offers a significant methodological advantage for imbalanced crash severity modeling.
• Empirical, data-driven optimization of synthetic data ratios is crucial, rather than fixed balancing strategies.
• The findings emphasize the importance of advanced techniques for improving traffic safety predictions.