Mathematics and Statistical Sciences Department.
Jackson State University. Academic Year: 2017-2018.
Departmental Colloquia & Seminars.
Dear colleagues,
The Departmental Colloquium Committee invites you to attend its 2017-2018 academic year’s activities on Friday 16, February 2018 at 2 :00 pm, room JSH 103.
Our speaker is Li-jing Arthur Chang a professor at Journalism and Media Studies.
Title:
Mississippi Traffic Accident Analysis: Comparison of Machine Learning Models for Car Riders Age 65 or Over
Abstract:
As many people are commuting daily to their work daily in America, traffic safety on the road is among the major concerns to both the commuters and traffic safety authorities. Given the huge amount of the traffic data collected over the years, it is possible to use the machine learning (ML) algorithms to identity major factors behind traffic accident severity (i.e., casualty or damage only) to predict the outcomes of future traffic accidents. Based on past research, the major factors contributing to traffic accident severity include (1) personal factors such as age, gender and alcohol use; (2) environmental factors such as weekdays, street-light conditions, weather conditions and accident time; and (3) accident factors such as rear-end collision, head-on and sideswipe. This study uses these factors and other related variables as the input features to predict the traffic accident severity outcome.
Based on the above empirical evidence, the following research question is generated: Which personal, environmental and accident factors have the most effect on casualty (fatality or injury) for people aged 65 and above? To test the research question, the study uses a dataset from the Mississippi Department of Transportation. It contains all variables mentioned above and other related features as well as the target variable, traffic accident severity.
The results of comparing accuracy rates among five ML models shows that Decision Tree is the best performing model, followed by Logistic Regression, SVM, Bayes Classification, and Association Rule. This means for this sample and among the five ML models considered, Decision Tree is the best model to predict target variable security (casualty or damage only) for car riders age 65 and over. By observing the improved (pruned) Decision Tree diagram, we can derive a few rules to help predict the severity target variable. For example, if damage extent is not heavy (Damage <= 2.5), the model will predict damage only. Follow similar logic and based on the tree, if the damage extent is heavy (Damage >= 2.5) and there is no sideswipe (Sidewipe < 0.5), the model also predicts damage only.