An Artificial Neural Network Approach for Road Accident Severity Prediction
Jameel Ahmad Sowdagur, B. Tawheeda. B. Rozbully-Sowdagur, Geerish Suddul
Abstract
Determining the underlying causes and severity of road accidents is a major concern worldwide. Several techniques have been used over the years, but more recently, Machine Learning (ML) approaches, especially Artificial Neural Networks (ANN), seems to be crucial. This research work focuses on the prediction of accident severity in the island of Mauritius. We experimented with different configurations of the multi-layer perceptron (MLP). The optimum values for hyperparameters were determined through systematic manual tuning over several experiments. They are as follows: 50 and 25 neurons in the hidden layers respectively, with a learning rate of 0.1, the Rectified Linear Unit (RELU) as an activation function and a maximum of 10,000 iterations. To test and avoid overfitting/underfitting, stratified 10-fold cross-validation was used. Our comparative analysis shows that the MLP outperformed all the other models previously implemented using the same dataset, with an accuracy of around 84.1%.