Travel Mode Detection Using GPS Data and Socioeconomic Attributes Based on a Random Forest Classifier

Abstract

The past few years have witnessed the rapid growth in the collection of large-scale GPS data via smartphone-based travel surveys around the world, following which transportation modes detection received significant attention. A mass of methods varying from Criteria-based rules to Machine Learning technology were employed to recognize the travel modes. However, the limited sample size, deficient feature selection and the less emphasis on addressing confusion modes, which leave room for improvement. This paper therefore sought to develop and evaluate a Random Forest classifier combined with a rule-based method to detect six travel modes (subway, walking, bicycle, e-bike, bus and car). Seven GPS-related variables are selected as feature set from the initial list of 22 variables. Consequently, more than 98% subway trips were correctly identified and the overall accuracy of the rest five modes classification is obtained as high as 93.11%. More than 85% trips were successfully identified for each mode except for the bus. More importantly, results show that socioeconomic attributes data could significantly improve the prediction of e-bike and address the confusion between bus and car modes. The employment of ROC curve provides a statistical proof to the excellent classification capacity of Random Forest in this study. Besides, the comparison with two representative classifiers demonstrates the applicability of Random Forest classifier for travel modes detection incorporating multi-source attributes.