8%) but performs worse on other modes. The rough sets model outperforms the prediction for the foot, bicycle, transit, and car modes. Another indicator, mean absolute percentage error (MAPE), was utilized to compare the coverage. MAPE is expressed as follows: MAPE=∑i=1nPEin,PEi=Xi−FiXi, (7) where PEi is the Hedgehog Pathway prediction percentage error of observations for the ith travel mode, Xi is the actual number of observations for the ith mode, and Fi is the predicted number of observations for the ith mode. The MAPE for rough

sets model and MNL model is 20.6% and 21.7%, respectively. Thus, the rough sets model proves to be better on the overall prediction coverage. 7. Conclusions This paper has demonstrated the successful application of a relatively new technique in the area of knowledge discovery to the well-studied problem of understanding and predicting traveler’s mode choices. The method has been able to reveal information about the household characteristics, individual demographics, and travel attributes with mode choices in a readily understandable form (a set of “IF-THEN” statements) and to use this information to predict mode choice for previously unseen individuals. The rough sets model shows high robustness of the model structure

to the training dataset due to their data induction property. No statistical assumptions (e.g., IIA property assumption) need to be made so the compatibility between the model structure and the observations is enhanced in the model estimation and hence the prediction performance can be improved. According to presence of derived rules, the most significant condition attributes

identified by the rough sets model of determining travel mode choices are gender, distance, household annual income, and occupation. Comparative evaluation with the MNL model shows that the rough sets model has comparable but slightly better prediction capability on travel mode choice modeling. The prediction results based on separate testing dataset show, on both accuracy and coverage, that the rough sets mode outperforms the MNL model. However, the rough sets induce too many detailed rules. Although the single rule is easy to interpret, the complete rule set is far too large to gain sound insight in travel behavior. Techniques such as generalization or shortening of the rule have been applied to deal with the problem [26]. Advanced models such as rough sets combined with genetic programming [30] can also be adopted in Anacetrapib the future to improve the performance of rule extraction and observations validation. Acknowledgments This research is sponsored by the National Natural Science Foundation of China (51178109) and the National Basic Research 973 Program (2012CB725402) and Chinese Postdoctoral Fund (2013M540408). The authors also would like to thank the graduate research assistants at School of Transportation at Southeast University for their assistance in data collection.