New Optimization Approach for Handling Imbalanced Data in Road Crash Severity

Accidents are a major problem that claim the lives of many people in the world each year. Fatalities and severe injuries could leave adverse and irreversible impacts on public health and economic prospects. A review of the variables affecting the severity of crash injuries can help reduce fatal accidents. However, a detailed prediction of fatal crashes as a smaller-data class than other classes is seen as a challenge. This study uses three robust machine learning such as Bayesian classifier, random forest, and support vector machine techniques. First, three imbalanced data prediction models were developed, suggesting they could not differentiate fatal data from injury data. To address this problem, three random, k-means clustering, meta-heuristic algorithms clustering techniques were used to balance the data. It should be noted that the genetic algorithm performed better than the particles swarm. Models developed by intelligent optimization methods, k-means clustering, and random methods were found to be more accurate, respectively. These criteria helped evaluate the models developed, which yielded the best model. The support vector machine method for genetic clustering-balanced data could predict fatal, and injury crashes with a 0.96% accuracy, becoming the best model. Finally, sensitivity analysis was performed on the best model, indicating that the highway, horizontal curves, and head-on variables contributed to fatal accidents.