Evaluation of the Performance of Random Forests Technique in Predicting the Severity of Road Traffic Accidents

Abstract

Traffic accidents in the Middle East are a primary concern for governments and local communities owing to the large numbers of fatalities, injuries and economic losses. Many analytical methods have been used in the literature to analyze the accidents database. One of the recent methods in this domain is the data-mining techniques. In this paper, we evaluate the performance of a well-known data mining technique called Random Forests (RF) in predicting the severity of road accidents based on 5973 accidents occurred in Abu Dhabi over a period of 6 years (2008–2013). The factors studied in this paper include: five accident-related attributes (year, day, time, reason of accident, and accident type), six driver-related attributes (gender, nationality, age, seat belt use, casualty status, degree of injury), and five road-related attributes (lighting, road surface, speed limit, lane numbers, and weather). The severity of the accident was classified into one of four classes (Minor, Moderate, Severe, and Death). RF was then used to build a prediction model using 10-fold cross validation method. The overall model predication performance was 68.5%. The generated model was found to perform poorly on the underrepresented classes (Death and Severe). As a result, the original data was transformed into a balanced data set using Minority Oversampling Technique (SMOTE). The performance of RF on the balanced data was 78.19% with 14% improvement. In order to validate the performance of the RF model, an ordered probit model was also used as a comparative benchmark. The accuracy of the ordered probit model was 59.5%, and 34% for the original and balanced data sets respectively. It was obvious that RF technique outperforms the ordered probit method in predicting the severity or road traffic accidents.