Extracting bus transit boarding stop information using smart card transaction data

As smart card (SC) fare system has been widely implemented in the world today, smart card data (SCD) plays an important role in the regional transportation system management. SCD can act as the data source to replace traditional travel surveys in many respects. Specifically, the transit rider origin–destination (OD) matrix can potentially be extracted from the SC transaction database. Passengers’ OD matrix estimation is a major part of any transportation planning study. With the consideration of inconvenience of traditional travel survey methods, SCD-based research has become more and more popular recently. The applications of SCD study are very broad. For example, Cui [1] utilized iterative proportional fitting and maximum likelihood estimation for both single route and network level OD estimation. Devillaine et al. [2] presented a methodology to categorize trips by different purposes after the detection and estimation of the locations of destination, trip time, activity duration and card type by utilizing smart card databases. Kusakabe and Asakura [3] developed a data fusion method to describe and add behavioral attributes to the smart card data by using survey-based data. By applying the probability distribution of behavior attributes to naive Bayes classifier, the attribute of trip purposes (‘commuting to work’; ‘commuting to school’; ‘leisure’; ‘business’; ‘returning home’) is added to each trip in the smart card database. Long et al. [4] conducted an analysis based on the travel time characteristics of four major passenger groups (‘early birds,’ ‘night owls,’ ‘tireless itinerants’ and ‘recurring itinerants’) with the help of household survey data. Kieu et al. [5] presented a study utilizing three levels of density-based spatial clustering of application with noise (DBSCAN) algorithm for mining transit users’ travel regularity (spatial regular OD and temporal habitual traveling time) from SCD. Ma et al. [6] developed an effective data-mining methodology utilizing Beijing transit data and also applied DBSCAN algorithm to analyze and detect transit riders’ historical travel patterns.

Although smart card fare systems are essentially different in each city and the relevant data components are unique in each case, in general, the first step of SCD-based OD estimation is to gather the boarding and alighting information. It is also important to note that in many transit systems, the smart card system is entry-swipe only which may not even have the information about boarding stops/locations. As such, different research methods need to be developed and applied for building up the OD matrix. In this regard, a number of SCD studies have been conducted to infer the OD matrix based on two important and yet reasonable assumptions as pointed out by Barry et al. [7]: (1) 'Most riders would like to return the destination station of their last trip to begin next trip; and (2) The destination of most riders at the end of day is the origin stop where they begin their first trip of the day.’ Based on these assumptions, inter-transaction time (ITT) was considered as a key element in transfer activities detection and OD estimation by Alsger et al. [8]. As for the buffer zone range calibration, different allowable transfer times (i.e., the threshold of ITT) were assumed which ranged from 30 to 90 min [6, 9‐11].

As mentioned, the trip origin detection is usually the first step of OD matrix inference when working with incomplete database. Association of the GPS data with smart card transaction data is one of the methodologies that can be used to infer boarding stops. Wang et al. [12] inferred the trip origins with the help of both transaction record timestamp and AVL data, and achieved some results with high accuracy. To solve the origin inference problem without the AVL system, Barry et al. [13] developed an approach to identifying the specific boarding stops utilizing scheduled run times to estimate the location of a bus along its route at the time of the AFC transaction. The challenge associated with the location estimation in their study was that the transaction times of MetroCard system were truncated to 6-min intervals. To detect the approximate bus boarding locations, they used scheduled run time between stops along the route and adjusted the results with the transfer information obtained from SCD.

Station labeling algorithm was used to assign each record of a bus route in a sequence. Song [14] utilized a fixed time threshold with an adjustment element to distinguish the boarding records at different stops with the time difference between adjacent records. This is the basic idea of boarding stops location detection. However, there are two problems in this algorithm that may contribute to errors. As known, the operation times between different stops can be different due to varying distances, number of changing boarding passengers and different signalized intersections involved. The specific time threshold between each pair of stops may result in the inaccuracy of stops labeling. Another problem is the consideration of the bus status at each terminal. One cannot be certain as to how long the bus stay will be in the terminal. Hence, using the only criterion ‘change transaction interval larger than bus dispatching time at terminus to detect direction change’ may not be accurate enough.

With only the transaction time being available, a few research efforts also attempted to adopt a hierarchical clustering method to classify the source data based on the time interval between two adjacent transaction records [15]. The reference bus operation time was based on the fixed schedule and the criterion used was based on the difference between timestamp of the cluster and the arrival time of each stop on the schedule. However, it was very difficult to solve the problem since the clusters could not be built without direction information.

Another approach was presented by Ma et al. [16], in which a Markov chain-based Bayesian decision tree algorithm was developed and utilized to extract passengers’ origin information from the Beijing flat-bus AFC database. The first task was to introduce how to cluster transaction data by taking the following two steps: (1) Sort transaction timestamps in ascending order; and (2) Include the record as another cluster if the time difference between itself and the previous record was larger than 60 s in which 30 min was set as the time threshold for a different trip. However, since the irregularity of the sample database and the time threshold configuration were a big issue, direction identification task was not described in detail in this research. In short, the following Table 1 part (a) provides a summary of the existing smart card research efforts for OD estimation and Table 1 part (b) gives a review summary of the existing smart card data research for origin inference.

The purpose of this research is to develop a systematic approach to illustrating how passenger boarding information can be mined only from the raw data derived from smart card-based automated fare collection (AFC) system, without the reference of GPS related data, AVL system, travel survey and other supporting database. In other words, there is only transaction time recorded, and no boarding stops locations and no bus operation direction information are available in this research. Advanced methods are developed to acquire the bus transit operation direction information, passengers boarding time information and the specific boarding stops information for each record. The methodology developed is tested via Guangzhou bus transit smart card data. It is important to note that effectively extracting boarding location information can greatly help build up the trip chain and also determine the destination information. The OD matrix can then be developed and used to help decision-makers plan, design, operate, and manage a more efficient public transit system. The contributions of this study include developing a systematic methodology which can be applied to most of the SCD systems and generating the OD estimation results which can reveal bus transit users’ travel pattern based on the collected SCD, both of which will greatly help transit operators and planning organizations develop objective and effective data-supported policies.

The remainder of this paper is organized as follows: Sect. 2 presents the methodology. The structure and potential problem of the smart card data are discussed. The direction identification algorithm is then described. This is followed by the boarding activities clustering and boarding stops identification. Section 3 discusses the numerical results of this study in detail. Finally, conclusions of the study are made, and the future research directions are also given in Sect. 4.

This section describes how to estimate boarding location information in detail. It first describes how to supplement the data when there is missing information about the transit directions and boarding clusters. Second, this section discusses the relationship between records in the sample database and presents the description of the boarding location inference methodology.

With the application of the algorithm as developed above to process the smart card data collected from Guangzhou, China, the boarding information is identified successfully which is presented in Table 4. There is a total of 100,000 smart card transactions with 98,632 of them being error-free. An example of such errors can be described as the consecutive swiping activity (e.g., transaction records of one specific smart card ID occurred more than three times in a same bus trip). By analyzing the raw data, information gathered out of this study includes:

Table 4 shows an example of the clustering results.

It is very challenging for one to conduct the OD estimation if there is no boarding stop information recorded by the AFC system. This research aims to develop a methodology to extract the boarding location information using the available transaction records with only basic route information, transaction time and few transfer activities, but without GPS, passenger counts, and survey data. To reduce the errors, the algorithm processes the data in the order of identification of boarding direction, boarding clusters and boarding stops. The boarding stop information is successfully extracted. Based on the acquired boarding information, the results of passenger counts are also derived in which the crest values of boarding activity occur during different peak hours. During the AM period, the highest volume occurs at 8–9 a.m. and the top three periods are 8–9, 10–11 and 9–10 a.m. During the PM period, the highest volume occurs at 5–6 p.m. and the top three periods are 5–6, 6–7, and 4–5 p.m. Most boarding activities occur at the first several stops, and the passenger boarding counts decrease as the bus stops get closer to the terminal.

However, several improvements could be made in the future, which include the consideration of variable speeds of the buses as one of the potentially influencing factors to calculate the operation time between bus stops. The direction changing time threshold of 30 min could also be further investigated as the fixed time threshold is utilized in this study. It is well noted that this time threshold could be changed during different times of day or days of week if the more transit data are available in the future.

The methodology and results of this study can be helpful for the OD estimation related work in the real world. However, with the limited amount of SCD, the trip chain building up process and destination identification framework are not discussed in this study. In the future, the transfer activities could be mined from the database if it contains information for more routes and longer periods. The cluster analysis can also be conducted to reveal passengers’ travel patterns with the help of survey data, card type, and land use data. Furthermore, the alighting stop information can also be extracted and used to estimate the relevant OD matrix.