Among discrete choice problems, travel mode choice modeling has received the most attention in the travel behavior literature. Most traditional mode choice models are based on the principle of random utility maximization derived from econometric theory. Alternatively, mode choice modeling can be regarded as a pattern recognition problem in which multiple human behavioral patterns reflected by explanatory variables determine the choices between alternatives or classes. The capability and performance of two emerging pattern recognition data mining methods, decision trees (DT) and neural networks (NN), for work travel mode choice modeling were investigated. Models based on these two techniques are specified, estimated, and comparatively evaluated with a traditional multinomial logit (MNL) model. For comparison, a unique three-layer formulation of the MNL model is presented. The similarities and differences of the models' mechanisms and structures are identified, and the mechanisms and structures in the models' specifications and estimations are compared. Two performance measures, individual prediction rate and aggregate prediction rate, which represent the prediction accuracies for individual and mode aggregate levels, respectively, were applied to evaluate and compare the performances of the models. Diary data sets from the San Francisco, California, Bay Area Travel Survey 2000 were used for model estimation and evaluation. The prediction results show that the two data mining models offer comparable but slightly better performances than the MNL model in terms of the modeling results, while the DT model demonstrates the highest estimation efficiency and most explicit interpretability, and the NN model gives a superior prediction performance in most cases.
Travel mode choice modeling from cross-sectional survey and panel data: the inclusion of initial nonresponse
