Artificial Intelligence Techniques for Predicting Individuals’ Mode Choice Behavior in Mansoura City, Egypt

Abstract

Predicting the choice behavior of individuals is an important step in transportation planning. This study examines the application of artificial intelligence (AI) techniques using machine learning with 10 different classifiers and deep neural networks (DNNs) to predict the travel mode choices of individuals in the city of Mansoura, Egypt, and compares the results with the traditional multinomial logit technique. The data used in this analysis were divided into two sets, training and testing, with a ratio of 67:33. The training dataset contains 10,173 cases, while the testing dataset represents 5,083 cases. The variables used to model mode choice behavior are: total travel time, total travel cost, gender, car ownership, driving license, occupational status, residency, and monthly personal income. The performance of all models is measured on two levels: the individual level, which predicts the overall percent of correct observations, and the aggregate level, which predicts the market shares for each mode. The results show that the DNN model, which takes mode captivity into account, outperforms all the examined models with a prediction accuracy of 97.81%. Additionally, random forest (RF), decision tree (DT), gradient booster (GB), and XGB showed prediction accuracy of over 95%. In addition, all the models were calibrated with different sample sizes. Prediction accuracy increases with increasing sample size, except for the Adaboost classifiers and linear discriminant analysis. With increase in sample size, the prediction accuracy of the RF, DT, GB, and XGB classifiers increases slightly while the DNN prediction accuracy increases significantly.