Machine learning approach versus probabilistic approach to model the departure time of non-mandatory trips

Abstract

Unbalanced distribution of trips during the time is one of the factors influencing traffic congestion at some hours of a day. Identifying the significant factors on travelers’ departure time choice and predicting their behavior helps maintain the balance in the time distribution of trips. For this purpose, this study employs and compares two machine learning and probabilistic approaches to model the departure time choice, including four choices, morning peak, noon peak, evening peak, and non-peak hours. Probabilistic support vector machine (PSVM) and multinomial logit (MNL) models calibrated based on the origin–destination data of Qazvin, and the evaluation and comparison of these two models made based on two applications of identifying the significant factors on the departure time and predicting the departure time. In terms of interpretability, the MNL model results have an indisputable advantage due to the lack of interpretable coefficients and parameters in the PSVM model. On the other hand, machine learning models’ predictive power partially covers the disadvantage of not being interpretable. The results show that the PSVM model can predict the departure time with 53.96% accuracy than the 49.98% accuracy of the MNL model. The maximum balanced accuracy for predicting morning peak, noon peak, and non-peak options is 69%, 53%, and 60%, respectively; obtained by the PSVM model and the MNL model predicts the evening peak option with a balanced accuracy of 52% more accurate than PSVM.