Prediction model of crash severity in imbalanced dataset using data leveling methods and metaheuristic optimization algorithms

Road accident is one of the important problems in the world which caused large number of deaths. In a road crash dataset, the fatal crash samples, often constitute very small proportion in comparison with non-fatal crash samples. Accurate prediction of fatal crashes, as a minority class, is one of the important challenges in such imbalanced sample distribution in the most of machine learning algorithms. This study introduced data leveling methods based on two metaheuristic optimization algorithms (biogeography-based optimization and invasive weed optimization) to obtain more balanced data. Then, three machine learning algorithms including decision tree, support vector machine (SVM) and k-nearest neighbor were applied for obtained balanced dataset. Performances of the prepared models were evaluated by improving the accuracy of the models in detecting the fatal crashes. It is found that data leveling methods of imbalanced dataset with metaheuristic algorithms improve the performance of crash prediction models in detecting fatal crashes especially in SVM algorithm up to 100% compared to previous studies. Also, results of sensitivity analysis on the developed model represented that head-on crashes, curved roads, and large type vehicles can increase the probability of fatal crashes up to 27.2%, 29%, and 36.8% at high posted speed limit, respectively. Also, two-vehicle crashes are much more likely to be involved in fatal crashes than single-vehicle crashes.