Assessing the Severity Level of Road Traffic Accidents Based on Machine Learning Techniques

摘要

Recently, predicting the road accident severity has attracted great attention since timely prediction is important to minimize the damage by the accidents. The problem we address in this study is to predict the severity of road accidents based on various machine learning techniques,such as decision trees, artificial neural networks, Bayesian networks, support vector machines (SVMs), and regression models. We also present the comparative analysis among those machine learning techniques. Our classification models are developed by employing the accident-related input parametersbased on 25,374 accident records occurred in Seoul, Korea over a 3-year period (from 2009 to 2011). For each accident, 30 attributes were collected at the time of accident. Through data preprocessing, our data were reduced to 12 attributes and three accident severity classes (Possible injuryaccident, Incapacitating and non-incapacitating evident injury accident, and Fatal accident). In this study, IBM SPSS Modeler that is the commercial data-mining S/W was used to build our classification models. The accident data were split into two disjoint sets, a training set (66.7%) forconstructing the models and a test set (33.3%) for validating the models. The experimental results revealed that SVM outperforms others in terms of prediction accuracy, and other classification models also achieved the considerable classification accuracy.