A MACHINE LEARNING MODEL FOR ROAD TRAFFIC ACCIDENT FACTOR CLASSIFICATION: THE CASE OF EAST GOJJAM ZONE

Abstract

Road traffic accidents are accidents that occur on a street open to public traffic. They result in many people's deaths or injuries, many vehicles' crashes, and damage of property. Most countries around the world, including Ethiopia, have increased the number of vehicles on the road. This has led to an increase in the number of road traffic accidents. Yet, it is regularly challenging to determine which specific conditions lead to such accidents. Various studies have been conducted to classify the severity levels of road traffic accidents, and others are focusing on whether the accident will occur or not. Moreover, most of the studies conducted so far focused on accident severity level prediction with fewer number of features. The study aims to classify the factors pertinent to road traffic accidents such as environmental factors, human factors, road factors, and vehicle factors m based on the independent variables collected from real road traffic accidents. The data for this study was collected from East Gojjam zone police office. It covers eight years accident record from the year 2006-2013 E.C. the preprocessed dataset has 19 attribute and 5250 instance. This study has been conducted using an experimental approach to determine the best-performing model. This study used the Python programming language for implementation and analysis purposes with different libraries. For this study, the researchers trained four different models, including Random Forest (RF), Decision Tree (DT), Multilayer Perceptron (MLP), and Extreme Gradient Boost (XGB) algorithm. Those models are selected based on an exhaustive study conducted to select the best performing model. In this study, evaluation of the model was done using percentage split (70/30), and classification performance metrics was used in order to compare the models. The finding of this study shows that the random forest classifier outperforms the rest of the classifiers with an accuracy of 98.7% on training data and 92.6% on test data. Based on this result, a system prototype was developed and tested that is capable of classifying factors of road traffic accidents.