< Back to previous page

Publication

Mobility Patterns for Context-aware Ridesharing Services

Book - Dissertation

Dynamic ridesharing is a mobility service where car drivers offer available seats in their car for sharing with passengers whose origin and destination happen to be close to their traveled path. It has gained interest as a potentially sustainable form of mobility, because it offers extra mobility by merely exploiting unused capacity that is otherwise lost, and hence at little to no marginal social cost - this in contrast to other alternatives for the private car like ridehailing services or public transport. Dynamic ridesharing however consist of a two-sided market, giving rise to a chicken-and-egg problem: without sufficient demand for trips, there is little incentive for a driver to offer his rides for sharing; without sufficient rides being supplied, there is little opportunity for passengers to choose the service. In order to attain a critical mass to guarantee the operability of dynamic ridesharing, one option is to train context-aware software agents to recognize the driver's mobility patterns and automatically predict and offer upcoming rides for sharing. To achieve this, this thesis develops an architecture of such a ridesharing app and the workflow of mobility history data collection and processing functions it should perform. It then develops enhanced iterative methods to identify in travel histories the user's personal points of interest, typical arrival times or transitions between locations and other mobility patterns that could be exploited to anticipate a ride. Without constraining ourselves to specific prediction techniques, this thesis assumes that in general, the predictive performance of any learning method depends on the regularity and frequency of patterns in the travel history (among others), and it proposes and tests novel methods to extract these multi-day characteristics from empirical life-logging data. A key asset to understand whether the set of all automatically shared trips in a region of interest would form an attractive supply for candidate ridesharing travellers, is simulation in various scenarios of the trips made and shared by a synthetic population. However, in existing synthetic travel demand, multiday characteristics of the trips are lacking, and hence one cannot determine which subset of all trips would be predictable and shareable by an automated ridesharing agent. To remedy this, the thesis proposes a method for generating synthetic multi-day trip demand for an urban region by merging through a newly developed form of statistical matching two complementary datasets: a set consisting of multi-day tracks from lifelogging data, and a second one with synthetic home tours for the same region network of Antwerp, Belgium. For successful matching, the intersection of both data sets should contain enough coincident characteristics; however, some of these matching features can only be discovered through appropriate data mining procedures. Through machine learning functions, describing the correlations that exist in the donor data between the matching features and the multiday characteristics, the missing multiday information was successfully transferred to the receptor database. Combined, the methodologies described in this research provide a way, based on the analysis of big data collected by mobile devices, to study how dynamic ridesharing could act as a mature travel mode in transport planning, and herewith to obtain recommendations for shared-mobility systems.
Publication year:2020
Accessibility:Open