Bayes classifiers for imbalanced traffic accidents datasets.

Area of Science:

• Road safety
• Data science
• Machine learning

Background:

• Traffic accident datasets often exhibit class imbalance, with fewer instances of severe injuries compared to minor ones.
• This imbalance poses a significant challenge for classification algorithms, leading to misclassification of critical injury cases.

Purpose of the Study:

• To address data imbalance in traffic accident datasets using various balancing techniques.
• To compare the performance of different Bayesian classifiers in identifying factors affecting accident severity.
• To analyze historical traffic accident data from Jordan to understand injury severity.

Main Methods:

• Applied three data balancing techniques: under-sampling, oversampling, and a combination of both.
• Compared Averaged One-Dependence Estimators, Weightily Average One-Dependence Estimators, and Bayesian networks.
• Utilized traffic accident data from Jordan spanning 2009-2011.

Main Results:

• Balanced datasets, particularly those using oversampling, enhanced the accuracy of Bayesian networks in classifying accident severity.
• The misclassification of killed or severe injury instances was significantly reduced.
• Identified key factors contributing to severe injuries: number of vehicles, accident pattern, number of directions, accident type, lighting, surface condition, and speed limit.

Conclusions:

• Data balancing techniques, especially oversampling, combined with Bayesian networks, are effective in improving traffic accident injury severity classification.
• This study provides the first analysis of historical traffic accident data in Jordan using balancing techniques for injury severity.
• The identified factors offer valuable insights for developing targeted road safety interventions.