Estimating motorized travel mode choice using classifiers: An application for high-dimensional multicollinear data. (January 2017)

• Record Type: Journal Article
• Authors: Lindner, Anabele; Pitombo, Cira Souza; Cunha, André Luiz
• Is Part Of: Travel behaviour and society. Volume 6(2017)
• Page Start: 100
• Page End: 109
• Publication Date: 2017-01
• Subjects: Artificial Neural Networks -- Decision tree algorithms -- Travel mode choice -- Multicollinear data; Transportation -- Periodicals; Population geography -- Periodicals; 303.48305
• DOI: 10.1016/j.tbs.2016.08.003
• Languages: English
• ISSNs: 2214-367X

Abstract

Highlights: CT and ANN may overcome disadvantages featured in travel demand traditional modeling. Two different levels of aggregation were considered: TAZ and household-related. CT and ANN turn out not to be affected by multicollinear information and outliers. Generally, data mining yields greater accuracy than logit models for travel mode choice. The use of classifiers has been seen as a positive tool towards travel behavior. Abstract: Studies in the field of discrete choice analysis are crucial for transportation planning. Generally, travel demand models are based on the maximization of the random utility and straightforward mathematical functions, such as logit models. These assumptions lead to a continuous model that presents constraints concerning fitting the data. Artificial Neural Networks (ANN) and Classification Trees (CT) are classification techniques that can be applied to discrete choice models. These techniques can overcome some disadvantages of traditional modeling, especially the drawback of not being able to model high-dimensional multicollinear data. This research paper compares the performance of estimating motorized travel mode choice through ANN and CT with a binary logit in a multicollinear study case (aggregated and disaggregated covariates). The dataset refers to an Origin-Destination Survey carried out in São Paulo Metropolitan Area, Brazil in 2007. Classification techniques have shown a good ability to forecast (approximately 80% match rate), as well as to recognize travel behavior patterns. Furthermore, by using the classifier application, the most important covariates within all the datasets can be selected. These covariates can be related to households, as well as to Traffic Analysis Zones.