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

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