Effectiveness of resampling methods in coping with imbalanced crash data: Crash type analysis and predictive modeling

Area of Science:

• Traffic Safety
• Data Science
• Machine Learning

Background:

• Crash data analysis frequently encounters imbalanced datasets, where rare crash types are often the most severe and costly.
• Existing modeling approaches often overlook data imbalance, leading to biased predictions and ineffective countermeasures.
• Developing robust models to predict severe, infrequent crash types is crucial for enhancing traffic safety.

Purpose of the Study:

• To compare the effectiveness of various resampling techniques in improving crash type classification and prediction models.
• To evaluate the performance of machine learning and classical statistical models on imbalanced freeway crash data.
• To identify optimal strategies for managing imbalanced data in traffic safety analysis.

Main Methods:

• A mixed sampling approach combining cluster-based under-sampling with three over-sampling methods (random over-sampling, synthetic minority over-sampling, adaptive synthetic sampling).
• Evaluation of four classification models: three ensemble machine learning models (CatBoost, XGBoost, Random Forests) and one classic statistical model (Nested Logit).
• Analysis of model performance in classifying and predicting different freeway crash types using imbalanced datasets.

Main Results:

• All investigated resampling techniques consistently improved the performance of both machine learning and statistical models.
• Adaptive synthetic sampling demonstrated superior performance among the over-sampling methods, significantly enhancing minority crash type prediction.
• Adaptive synthetic sampling improved rare crash prediction without negatively impacting the prediction accuracy of common crash types.

Conclusions:

• Resampling techniques are essential for developing robust and unbiased models for imbalanced crash data.
• Adaptive synthetic sampling is a highly effective method for addressing data imbalance in traffic crash analysis.
• The findings provide valuable insights for developing improved traffic safety models and effective countermeasures for severe crash types.