Behavioural findings from observed transit route choice strategies in the farecard data of Brisbane
Behavioural findings from observed transit route choice strategies in the farecard data of Brisbane
Neema Nassir (corresponding),
Postdoctoral Research Fellow, University of Queensland, Australia
Mark Hickman,
Professor and Chair of Transport Engineering, University of Queensland, Australia
Zhen-Liang Ma,
PhD Student, University of Queensland, Australia
Abstract
This paper contributes to the behavioural understanding of transit passengers' route choice decisions in dense transit networks. We explore the route choice boarding strategies of passengers in the Brisbane area (QLD) by studying the farecard transactions in the Brisbane network.
In strategy-based transit demand forecasting models, public transit passengers are assumed to plan their travel based on a travel strategy that contains one set of attractive routes per boarding stop. Among these attractive routes, the passenger will board the route that arrives first to their location. For regular passengers of several high demand origin-destination (OD) pairs in the Brisbane area, we have inferred the actual attractive sets through a longitudinal analysis of the farecard transactions. In this paper, we present empirical findings related to these observed attractive sets. We explore variability of these sets among the passengers, and develop models to predict the attractiveness of a route. Based on these model outcomes, important attributes that can explain the passengers boarding behaviours are presented and discussed, and new behavioural aspects are uncovered.
1. Introduction
A considerable literature over the past fifty years, especially in transit path assignment, have identified and accounted for situations under which the passengers’ path considerations can vary from an elementary path between the Origin-Destination (OD) pair to a complicated and adjustable strategy that includes many possible elementary paths. A recent analysis performed on the farecard data from London (Oyster card) revealed that most commuters do not use fixed routes for their regular commutes (Kurauchi et al. 2014), supporting the importance of hyperpaths and strategies in transit path choice analysis. Fonzone et al. (2010, 2013) also reported a similar finding from a web-based survey that was distributed among international respondents from 106 cities in 25 countries (Fonzone et al. 2010, Fonzone et al. 2013).
Kurauchi et al. (2012) estimated a hyperpath choice model, with in-vehicle time, waiting time and number of transfers as utility attributes of the choice model. They used a web-based stated preference survey with hypothetical scenarios in an abstract network with three routes and up to seven hyperpaths. They asked their respondents which of the available hyperpaths they would choose if they had to travel in the hypothetical network. Their analysis demonstrated differences in the hyperpaths constructed by different socio-demographic groups and for different travel purposes. An advantage of a stated preference survey is that the information can be collected about the passenger strategy which is difficult (if not impossible) to infer from revealed preferences in a typical travel survey. However, a drawback of such a survey in abstract networks is whether the resulting choice models are realistic, especially in real-sized and dense transit networks. In the context of hyperpath behaviour in a real-sized network, (Fonzone and Bell, 2010) focused on human rationality and limitations on computational ability in making complex hyperpath choices. Considering these realistic limitations, they proposed a myopic hyperpath calculation method and included a limit on the size of the hyperpath that is handled by the travelers in a real-sized network.
The need to incorporate realistic behavioural aspects of strategy choices, especially in dense and complex transit networks, motivates new choice modelling methodologies based on real world travel observations. However, actual travel strategies are not easily observed with traditional data collection methods. For instance, in a travel survey, a record captures only one possible outcome of a (latent) strategy that has appeared under the particular conditions of the service (e.g., the order of bus or train arrivals) at the time and location of observation. For a full strategy, all possible instances of service conditions that affect the path outcomes should be enumerated, and travellers’ responses to each of those instances should be collected. (Raveau et al. 2014) performed a survey at one of the central transit stations in Santiago Chile and collected path choice strategy information along with socioeconomic information from the passengers departing that station. They classified the respondents into three strategy choice categories and developed a model to predict the strategy choice category for the passengers at the upper level, and the path choice in each category at the lower level. While this research can be a good start in strategy choice behavioural analysis, the requirement to collect full strategy information from larger samples can probably prohibit direct data collection approaches in network-wide studies. This gap can be filled by using alternative sources of travel data that are now available, such as transit farecard transactions.
For a farecard dataset from a local city in Japan, (Schmöcker et al. 2013) proposed a bi-level discrete choice structure to model the passengers’ hyperpath choices. They proposed a hierarchical model that assumes the choice of hyperpath is based on personal preferences at the higher level, and deterministic probabilities of boarding on the hyperpath routes are proportional to their frequencies on the lower level. Their model estimates hyperpath utility coefficients for attributes of waiting time and in-vehicle time.
Viggiano (2013) explored passenger boarding choices in two multi-route bus corridors in London, using Oyster farecard transactions. They proposed and tested probabilistic models to deduce aggregate and disaggregate (individual) strategic boarding behaviour of passengers. They classified the passenger strategies of the corridors travellers into two categories: “first-bus” strategy and “favourite-bus” strategy, and inferred the percentage of passengers belonging to each category. They explored and observed the effects of in-vehicle travel time, frequency of corridor use, crowding, and access to real-time countdown information on passenger choices. Viggiano et al. (2014) sent an online survey to the registered farecard passengers of the two corridors to enquire about their boarding strategies. Their survey results confirmed the strategy assumptions that were deduced from their farecard analysis.
Nassir et al. (2015a) developed a statistical inference algorithm to apply to longitudinal farecard transactions in order to infer a reliable estimate of passenger route choice attractive sets. Their algorithm deduces the “steady-state attractive sets” from passengers’ repeated route choice records between given origin-destination (OD) pairs, by detecting and eliminating the occasional choices and tried-and-failed path experiences. In this paper a case study is conducted and that algorithm is applied to 6 high demand OD pairs in the network of Brisbane, Australia, and the passengers’ route choice attractive sets are inferred. In this case study: 1) the variability of these attractive sets among the passengers is explored; and, more importantly 2) statistical models are developed to predict the probability that a typical passenger considers a given route in their attractive set, and interesting behavioural aspects are reported and discussed.
2. Review of strategy inference algorithm
In strategy-based transit assignment models, each passenger’s travel strategy contains one set of attractive routes per boarding stop (called attractive set), and it is assumed that the passengers will board the route in the attractive set that arrives first to the boarding stop. Nassir et al. (2015a) proposed an algorithm to infer these attractive sets by investigating the farecard records of each passenger over a long period of time. Since farecard datasets usually do not contain actual bus and passenger arrival times to the stop locations, Nassir et al. (2015a) proposed a probabilistic estimation of the attractive sets based on the assumption of schedule-based or frequency-based bus arrivals, and uniform (Poisson) passenger arrivals during the study period. They introduced the notion of “steady-state strategies” which refers to the set of attractive routes for each passenger of the given OD pair that is assumed to be used constantly and continuously over the study period. Based on this notion, unsteady-state patterns which may be the case during the learning process of the path choice, or due to occasional choices are either corrected or separated in the output.
The algorithm starts with the set of all recorded routes (Rn) that passenger n has boarded to travel between the OD pair over the study period, as a potential attractive set. Then, a statistical test is performed to check if the longitudinal boarding records of passenger n could be a random distribution of boardings onto the Rn routes (under random arrivals of buses and passengers to the stop). In other words, if the difference between the observed longitudinal boarding distribution and the expected random distribution is not statistically significant, Rn is inferred as the attractive set of routes. Otherwise, Rn is rejected and modified to find another potential attractive set. The modification is made by removing the route (and all boarding observations on that route) that has the lowest probability of being in the attractive set. Then, the same statistical test is performed again, and this continues until either an acceptable attractive set is inferred, or the number of observations that are removed becomes a considerable portion of the total observations, in which case the algorithm stops without inferring any attractive set.
3. Farecard data (Go Card)
The research reported in this paper is conducted using a farecard dataset from the public transit services in Southeast Queensland (SEQ) in Australia. The farecard data consists of six months of transactions of the regional farecard, the “Go Card”. The Go Card system records both the boarding and alighting locations, and the time of these transactions, for a clear majority (85%) of all passenger journeys in SEQ. Since farecard identifiers are available in the Go Card dataset, the longitudinal analysis of the path choices becomes possible. We apply the proposed method to a selected set of high demand origin-destination (OD) pairs in SEQ. Six months of Go Card data, from November 2012 to April 2013, is filtered for the set of “regular” passengers that travelled between 6 transit OD pairs during the weekday afternoon peak (4pm-6pm). “Regular” passengers are the ones who travelled these OD pairs during the evening peak 10 or more times in the six-month period.
When searching in the farecard dataset for travel trajectories, we include both direct and indirect (i.e., through transfer interchanges) paths in the study dataset. In order to include the indirect paths, we have applied a transfer identification algorithm that has been calibrated using the same farecard dataset (Nassir et al. 2015b). This algorithm can distinguish the transfer exchanges from intermediate activity locations using the spatio-temporal characteristics of the observed trajectories and the scheduled service.
4. Inferred boarding strategies
The origins and destinations of the case study OD pairs in this research are selected from high demand nodes in the Brisbane network. The origins are either at University of Queensland main campus (UQ), or at a major bus station located in the central business district (CBD) of Brisbane-- cultural centre bus station (CC). The destinations are selected from major shopping centres and residential suburbs that have a large inflow of passengers from UQ and CC during the evening peak. These destinations are Toowong (TW), Indooroopilly (IP), Garden City (GC), and Carindale (CD). The selected OD pairs between these locations are “UQ to TW”, “UQ to IP”, “UQ to GC”, “UQ to CD”, “CC to GC”, and “CC to CD”.
The number of regular passengers of these OD pairs range from 88 (travelling from CC to CD) up to 803 passengers (travelling from UQ to TW), and in total sum up to 2,453 passengers. By applying the strategy inference algorithm, the boarding strategy for 2,320 passengers (94.6%) is inferred. Strategy inference results are summarised and presented in Table 1.
In Table 1, the first column identifies the OD pair. The second column shows the inferred attractive sets, which are the set of routes between the OD that are constantly and continuously used by at least one passenger of the OD. It should be noted that the set of the attractive routes reported in Table 1 is inclusive, meaning that there is significant statistical evidence that the passenger with that attractive set has not boarded any other routes (regularly) than those in the attractive set, regardless of arrivals. The inferred attractive sets reported in Table 1 include both direct routes between the OD pair and routes that require at least one transfer interchange (transfer routes). Transfer routes in Table 1 are underlined to be distinguished from direct routes.
The third column in Table 1 shows the number of passengers inferred with every recorded boarding strategy (i.e. attractive set). Column 4, 5, and 6 respectively show the average number of trips recorded for the passengers with every recorded strategy, the mean value of α likelihood, and the mean value of α’ likelihood over all passengers inferred with each strategy. Probabilities α and α’ are the estimated likelihood measures of the inferred strategies for each passenger, based on a frequency-based assumption (α) or on a scheduled-based assumption (α’) about passengers boardings.
As it can be observed in Table 1, there is a considerable variability among the choice of attractive sets by regular passengers of every OD pair. The number of inferred strategies ranges between 4 (UQ to GC) and 24 (UQ to IP) different strategies.
Some general observations can be made about the strategy choice behaviour of passengers in all OD pairs. A large portion of passengers only choose direct routes; but there are also some passengers in all OD pairs who tend to board transfer routes too. Among the passengers that board only the direct routes, there is usually a considerable portion that uses all of the possible direct routes. However, when the number of direct routes serving the OD pair increases many passengers tend to act selectively and choose a subset of direct routes in their attractive set, and refuse boarding the rest of them.
In the next section we develop statistical models to realise the contributing factors that influence the choice of boarding a route or refusing to board it.
Table 1. Inferred boarding attractive sets