Using Machine Learning to Predict Injury Severity of Road Traffic Accidents During New Year Festivals From Thailand’s Open Government Data

The fast-rising number of vehicles on the road has resulted in a matching number of traffic accidents, especially during long holidays when people travel back home. A better understanding of the risk factors leading to traffic accidents may help decrease the number and prevent severe injuries and deaths. This paper addresses the imbalanced binary classification problem of predicting injury severity of road traffic accidents during Thailand's new year festivals using Thailand's open government data. Random undersampling, oversampling with SMOTE, and the combination of both undersampling and oversampling were used to rebalance training instances before fitting any Random Forests models. The results show that random undersampling gives the best performance at 83 percent with regards to both recall and accuracy. Important factors driving the prediction of severe injuries are delivering method, age, and alcohol drinking. Our work could benefit public authorities in analyzing accidents' hot spots and crucial factors that discourage road safety.