Modeling the destination choice has been of great interest for travel behavior community as well as policymakers in understanding the demand for land use and transportation infrastructures at aggregate and disaggregate levels and possibly devising policies to balance the demand and supplies. One of the challenges underlying predictions of location choice is the large choice set. While traditionally many methods had been devised to limit the choice set size either on a rather ad hoc basis or based on space–time prism by removing the locations out of reach of the subjects, the current study takes a substantially different approach and proposes a data-driven method to customize the generation of the choice set. The proposition is that observing the mobility patterns of citizens for multiple weeks would enable us to limit the choice set, depending on how far the subjects travel (beyond or within the distance they travel for their most frequent activities) to conduct their various activities. More precisely, using longitudinal trajectory data, we first classify people into two subgroups: returners and explorers, based on the size of the area (around their k most visited locations: k-radius of gyration) they move during the observation period. The destination choice set for four types of activities is then customized for returners (and explorers) and is used in a sequence of decisions represented by decision trees for the prediction of their destinations. The models for the whole sample and each subgroup separately are compared. The results suggest that the accuracy of destination prediction improves substantially for all four selected activity types, especially for the returners whose choice sets are formed based on their radius of gyration.
Read full abstract