Comparison of supervised ML algorithms for road traffic crash prediction models in Rwanda

Abstract

The development of Rwanda was accompanied by rapid growth of various modes of transport and the occurrence of road traffic crashes (RTCs). RTC severity has also increased in Rwanda, with the number of fatalities increasing progressively every year. The aim of this study was to compare the classification performance of eight supervised machine learning (ML) algorithms in order to determine the best one to predict crash severity and identify potential RTC-influencing factors in Rwanda. Quantitative data sets of RTCs, numbers of registered vehicles and annual average daily traffic (AADT) from 2010 to 2022 were used. The ML algorithms examined were logistic regression (LR), support vector machine (SVM), naive Bayes (NB), K-nearest neighbour (KNN), random forest (RF), decision table (DT), lazy Bayesian rules (LBR) and J48. Five algorithms (RF, DT, J48, LBR and KNN classifiers) were found to have an accuracy of more than 80%. The RF classifier was found to have the best performance for predicting crash severity in Rwanda, with an accuracy of more than 97%. The most influential factors were identified as AADT, number of registered vehicles, causes of crashes and the type of vehicles involved. The model results can be used to provide useful information to road safety decision makers during the planning and design of road infrastructure.