Modelling route choice behaviour with incomplete data: an application to the London Underground

This thesis develops a modelling framework for learning route choice behaviour of travellers on an underground railway system, with a major emphasis on the use of smart-card data. The motivation for this topic comes from two respects. On the one hand, in a metropolis, particularly those furnished with massive underground services (e.g. London, Beijing and Paris), severe passenger-traffic congestion may often occur, especially during rush hours. In order to support the public transport managers in taking actions that are more effective in smoothening the passenger flows, there is bound to be a need for better understanding of the passengers’ routing behaviour when they are travelling on such public transport networks. On the other hand, a wealth of travel data is nowadays readily obtainable, largely owing to the widespread implementation of automatic fare collection systems (AFC) as well as popularity of smart cards on the public transport. Nevertheless, a core limitation of such data is that the actual route-choice decisions taken by the passengers might not be available, especially when their journeys involve alternative routes and/or within-station interchanges. Mostly, the AFC systems (e.g. the Oyster system in London) record only data of passengers’ entry and exit, rather than their route choices. We are thus interested in whether it is possible to analytically infer the route-choice information based on the ‘incomplete’ data. Within the scope of this thesis, passengers’ single journeys are investigated on a station basis, where sufficiently large samples of the smart-card users’ travel records can be gained. With their journey time data being modelled by simple finite mixture distributions, Bayesian inference is applied to estimate posterior probabilities for each route that a given passenger might have chosen from all possible alternatives. We learn the route-choice probabilities of every individual passenger in any given sample, conditional on an observation of the passenger’s journey time. Further to this, the estimated posterior probabilities are also updated for each passenger, by taking into account additional information including their entry times as well as the timetables. To understand passengers’ actual route choice behaviour, we then make use of adapted discrete choice model, replacing the conventional dependent variable of actual route choices by the posterior choice probabilities for different possible outcomes. This proposed methodology is illustrated with seven case studies based in the area of central zone of the London Underground network, by using the Oyster smart-card data. Two standard mixture models, i.e. the probability distributions of Gaussian and log-normal mixtures, are tested, respectively. The outcome demonstrates a good performance of the mixture models. Moreover, relying on the updated choice probabilities in the estimation of a multinomial logit latent choice model, we show that we could estimate meaningful relative sensitivities to the travel times of different journey segments. This approach thus allows us to gain an insight into passengers’ route choice preferences even in the absence of observations of their actual chosen routes.