The development of a segmentation scheme for the evaluation and prediction of activity-travel sequences

Abstract

In this paper, sequential information in data is represented and captured through the use of Markov Chains. The core knowledge information in a Markov Chain is typically represented by means of transition matrices, revealing information about the underlying structure of the data sequence. A drawback of the current application of Markov Chains is that there is only one transition probability matrix which is both representative for every person (respondent) and for every time frame during the day. To this end, a novel segmentation procedure has been introduced and tested in this paper that enables one to cluster transition matrices in terms of time and socio-demographic information. The temporal segmentation used the technique of the identification of bifurcation points; the socio-demographic segmentation used a modified version of a decision tree, in the sense that sequential probability information was used during induction and in the leaves of the tree as opposed to the traditional way of only using one single classification attribute. The segmentation procedures were both adopted for descriptive and predictive purposes in the empirical section. Results show that the technique reveals promising information both at the descriptive and predictive level. At the descriptive level, evidence was found that one should rely upon different transition probability matrices for different time windows during the day and that socio-demographic information should be taken into account as well. For prediction purposes, the segmentation approaches simulated more accurate activity-travel sequences at pattern level; while the opposite was found true at trip level.