Predicting Accident Severity using Machine Learning

Abstract

Each year, millions of people are killed in automobile accidents. Predicting the severity of an occurrence allows local authorities to respond quickly and save many lives. Information and data on traffic accidents made available by public organizations can be used to categories these incidents based on their nature and severity, and then attempt to construct predictive models that can be further investigated to identify fatal accident risk factors. The study provides ways to develop a system for determining the severity of accidents. To evaluate the severity of the collision, we use a variety of weather-related characteristics such as temperature, humidity, visibility, and pressure, as well as several other circumstances, such as the presence of a traffic light or a junction. The Select-K-Best features selection algorithms were used to choose the best characteristics from a list of 47. The accuracy of both balanced and unbalanced data was measured, and the balanced data was chosen for further analysis. The balanced dataset is then trained, analyzed, and compared with a number of machine learning approaches, including the Random Forest Classifier (RFC), K closest neighbour (KNN) classifier, and Naive Bayes classifier. The RFC classifier outperformed with an AUC of 95%, whereas the naïve bayes underperformed. Furthermore, the accuracy is improved by using all of the aforementioned methods as base learners for stacking ensemble models with logistic regression as meta learner. This stacking ensemble strategy outperforms RFC with improved precision and accuracy, resulting in an AUC of 96.92%. The results revealed that the model has a 96.92% AUC in predicting accident severity. The work can be extended to investigate the complex correlations between a few key parameters and accident severity.