See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/339239174

On the tradeoff between sensitivity and specificity in bus bunching prediction

Article in Journal of Intelligent Transportation Systems Technology Planning and Operations · February 2020
DOI: 10.1080/15472450.2020.1725887

CITATIONS READS

7 389

3 authors:

Wenzhe Sun Jan-Dirk Schmöcker

Kyoto University Kyoto University
20 PUBLICATIONS 145 CITATIONS 193 PUBLICATIONS 3,236 CITATIONS

SEE PROFILE SEE PROFILE

Toshiyuki Nakamura
Nagoya University
46 PUBLICATIONS 249 CITATIONS

SEE PROFILE

All content following this page was uploaded by Wenzhe Sun on 25 February 2020.

The user has requested enhancement of the downloaded file.

Unformatted version of: Sun, W., Schmöcker, J. D., & Nakamura, T. (2020). On the
tradeoff between sensitivity and specificity in bus bunching prediction. Journal of
Intelligent Transportation Systems.

Available from: https://doi.org/10.1080/15472450.2020.1725887

On the trade-off between sensitivity and specificity in bus bunching prediction

Wenzhe Sun, Jan-Dirk Schmöcker, Toshiyuki Nakamura

Wenzhe Sun
Department of Urban Management, Kyoto University, Kyoto, Japan
orcid: 0000-0002-7305-8671
Email: [email protected]

Jan-Dirk Schmöcker, corresponding author

Department of Urban Management, Kyoto University, Kyoto, Japan
orcid: 0000-0003-2219-9447
Email: [email protected]

Toshiyuki Nakamura
Institute of Innovation for Future Society, Nagoya University, Nagoya, Japan
orcid: 0000-0003-3093-4229
Email: [email protected]
On the trade-off between sensitivity and specificity in bus bunching prediction

Abstract

Bus bunching resulting from initially small headway irregularities is a widely-known and studied
problem. A variety of headway-prediction approaches, as well as corrective strategies, have been
developed to identify and correct headway irregularity in real time. Instead of predicting an exact value
for future headways, this study explores a probabilistic predictive methodology to forecast whether or
not a bus will be bunched during its dwelling at a downstream stop, using a logistic regression model
based on GPS records of buses at least k stops upstream to allow for sufficient time to implement
control strategies. A case study is conducted on a circular bus route in Kyoto City. Compared to two
headway-based prediction approaches using linear regression and support vector machine, the superior
performance of the proposed tool in detecting bunching is illustrated by Receiver Operator
Characteristic (ROC) analysis. The high reliability in long-term prediction gives adequate time for
operators to employ countermeasures. Besides, the proposed method provides operators with trade-
off options. We find that a bunching-averse operator can obtain 95% “sensitivity”, that is the ratio of
correctly identified bunching events, at the cost of decreasing “specificity”, which is the ratio of correct
non-bunching predictions over all events. This is true even if the prediction horizon is more than 10
stops.

Keywords: bus bunching prediction; logistic regression; sensitivity and specificity; bus GPS data;
multiple-stop-ahead prediction
1. Introduction

Bus bunching is a frequently occurring undesired event. Generally it can be defined as the phenomenon
of two successive bus runs of a single line arriving at a stop within significantly shorter headways than
the designed one. Bunching involving more than two buses is also regularly observed. Bus bunching
may be initiated by the arrival of one bus run being delayed at an upstream stop. More passengers are
likely to accumulate for the delayed bus at that stop and the bus is thus further delayed. Conversely,
the subsequent run has fewer passengers to pick up and departs earlier than scheduled. Accumulated
delay to the first vehicle and increasingly earlier arrival of the second one result in obvious inequality
in dwell times and on-board passenger numbers. As the inequality aggravates over a sequence of stops,
the scheduled headway is significantly shortened or eventually offset and the leading bus among
bunched bus is often overcrowded.
Accurate prediction on headway or bunching itself can help to spotlight the coming bunching and
further assist the operator to eliminate bunching in real time. A useful prediction tool is expected to a)
have a long enough prediction horizon to allow the operator’s implementation of countermeasures and
b) provide information on the reliability of the prediction. The latter point is important in order to
account for different preferences among operators. A bunching-averse operator is willing to frequently
control the service to avoid any possible bunching, whereas some other operators may hesitate to take
control action that will negatively impact some passengers, they thus only correct the predicted
bunching of high confidence level. Therefore, this paper suggests a probabilistic binary prediction
method.
This study aims to extend the existing literature in two aspects. Firstly, this study builds a logistic
regression (LOGR) model to predict the likelihood of bunching to occur using bus GPS data, and tests
the prediction performance under a wide range of prediction horizons varying from 1-stop-ahead to
15-stop-ahead, with an emphasis on multi-stop-ahead prediction and understanding the regularity
deterioration pattern. Secondly, this study tries to enhance the robustness and flexibility of existing
prediction tools. To achieve this Receiver Operator Characteristic (ROC) curves are utilized. This
method is widely used in evaluating the performance of binary classification models and in this study
it is interpreted as the optimal front of the proposed LOGR. This study explains how to conduct the
trade-off between “sensitivity” and “specificity” from an operator’s perspective.
The paper is organized as follows. After this introduction, Section 2 conducts a literature review on
the corrective and predictive models addressing the bus bunching problem. The predictive
methodology using LOGR is elaborated in Section 3. We point out that LOGR might be biased when
used for “rare events data” as is the case in our example and provide a correction method. Then two
headway-predicting algorithms: linear regression (LR) and support vector machine (SVM) are taken
as the two benchmark approaches in this study and are also briefly introduced in this section. In Section
4, the characteristics of the collected data are described, including data collection period, average stop-
to-stop travel time, average scheduled headway, fluctuation patterns for headway, etc. Based on this,
a proper prediction horizon and bunching threshold are determined. The case study is described in
Sections 5-7. In Section 5, the prediction performance of the two headway-predicting algorithms is
discussed. The prediction performance of the proposed LOGR is evaluated and compared with
headway-based methods in Section 6. The trade-off functionality of LOGR is discussed in Section 7.
Conclusions and further work can be found in Section 8.

2. Literature review

Most of the relevant existing literature can be cast into two categories according to their objective:
bunching prediction and corrective strategies. A large body of literature discussed how to eliminate
bus bunching using analytical or simulation methods following the seminal work by Newell and Potts
(1964). Osuna and Newell (1972) and Newell (1974) tried to maintain the bus schedule by a single
control point. On the other hand, advanced control methods such as dynamic holding control proposed
by Eberlein, Wilson, and Bernstein (2001), Daganzo (2009), Xuan, Argote, and Daganzo (2011),
Bartholdi and Eisenstein (2012), Zhang and Lo (2018) and velocity control developed by Daganzo
and Pilachowski (2011) as well as stop skipping discussed by Sun and Hickman (2005) assume
frequent and efficient communication between bus drivers and the control center. Berrebi et al. (2018)
tested the control strategies proposed by Dagazo (2009), Xuan et al. (2011), Bartholdi and Eisenstein
(2012), Daganzon and Pilachowski (2011), Berrebi, Watkins, and Laval (2015) on a bus route in
Portland, Oregon. The experiment was based on bus automatic vehicle location (AVL) data, automatic
passenger counter (APC) data and traffic signal data. The effectiveness of each strategy to stabilize
bus headways was confirmed. Further, the effect of incorrect future headway prediction on each
strategy was discussed. The variance of controlled headway was found rising significantly as the
prediction errors increased. Instead of actively adjusting the headway, Schmöcker, Sun, Fonzone, and
Liu (2016), Wu, Liu, and Jin (2017), Sun and Schmöcker (2018) discussed passive strategies such as
passenger re-distribution and overtaking which are activated when bunching occurs. These strategies
aim to equalize passenger boarding numbers for bunched buses through queue management.
Substantial development in data collection technology recently gives scholars access to massive bus
operation data including AVL, APC and automatic fare collection (AFC) data, and has led to a large
number of studies concerning real-time prediction of bus operational aspects. Rather than predicting
bus bunching events, most existing literature focuses on bus arrival time and headway. Though closely
related, this literature can again be grouped into three subcategories: bus trajectory, bus arrival time
and headway prediction. Complete bus trajectory prediction is most challenging but also most
informative. It provides predicted stop arrival and departure time, stop-to-stop travel time, as well as
the headway between consecutive buses for bus operators and users. Hans, Chiabaut, Leclercq, and
Bertini (2015) developed a sequential mesoscopic simulation that elaborately considered the
stochastics generated during bus dwell time and link travel time. A bundle of possible future
trajectories is simulated based on the distribution assumed for the time components in a bus trip and
the associated parameters are calibrated with AVL, APC and traffic signal data. This method delivering
robust prediction results to the operator. Distribution or range for future arrival time and headway can
also be easily obtained. A shortcoming of this method is that the predicted range of arrival time or
headway might be too wide to be conclusive for operators’ decision making. Recent research by Dai,
Ma, and Chen (2019) also modeled bus dwell time and link travel time in detail to reproduce the trip
travel time variability for a bus line. They specifically considered the bus waiting time due to the
interaction between buses at the stop intersected by multiple bus lines, which is also defined as
common-line bunching in Schmöcker et al. (2016). They inferred the probabilities of the bus from a
specific line queueing (bunching) after the other common lines at the stop from bus GPS data. Yu,
Chen, Wu, Ma, and Wang (2016) conducted a solid literature review on the methods addressing bus
arrival time prediction. They reviewed the implemented data source and algorithm of each relevant
literature. SVM, Kalman filter (KF), k-nearest neighbor (KNN), artificial neural network (ANN) and
regression-based methods are frequently used. Yu, Yang, and Yao (2006) made a successful attempt at
predicting bus arrival time based on SVM method and AVL data. Yu, Lam, and Tam (2011) used SVM,
ANN, KNN and LR to predict arrival time for a 0.7km common line section where more than 10 bus
routes overlapped in Hong Kong. Kumar, Vanajakshi, and Subramanian (2018) combined KF and
KNN to tackle the prediction of bus travel time and arrival time. In this hybrid model, KNN classifier
is used to refine the model input of KF model.
Future headway is the difference between the predicted arrival times of two consecutive buses and
can be obtained by the arrival time prediction model. There are also some studies directly focusing on
the prediction of headway itself. Yu, Wu, Chen, and Ma (2017) proposed a probabilistic prediction
approach using RVM (Relevance Vector Machine) to attach a confidence interval for each predicted
headway for 2- and 3-stop-ahead. Outperformance with respect to robustness was concluded by
comparing the results with the deterministic single values derived by SVM, KF, KNN and ANN
algorithms. Andres and Nair (2017) integrated headway prediction and bus holding control strategies.
Regression, ANN and autoregressive models are used in their work to predict future headways with
5min and 10min prediction horizons. The prediction results are applied as input to an analytical model
extending Daganzo (2009).
Although headway prediction methods have made great advancement, it remains a challenging work
to precisely identify coming bunching events in multiple-stop-ahead prediction. The accuracy of
bunching prediction is heavily dependent on the reliability of headway prediction whose results
deteriorate gradually as the prediction horizon extends. Yu, Chen, Wu, Ma, and Wang (2016) used
several well-developed algorithms to predict headway first then convert the result to binary bunching
occurrence. 2min RMSE is obtained for headway and 99% sensitivity is realized for bunching in 2-
stop-ahead prediction, but the performance deteriorates to 6min RMSE and 73% sensitivity for 5-stop-
ahead prediction. Moreira-Matias, Cats, Gama, Mendes-Moreira, and De Sousa (2016) built a
regression-based model to predict the headway for a downstream stop and calculate the likelihood of
bus bunching to occur for all the further downstream stops. The focus of their study was to propose a
proactive control framework in which every suspicious event triggers a bunching alarm. The effect of
bunching likelihood thresholds was not investigated. It should be noted that Moreira-Matias et al.
(2016), Andres and Nair (2017), Berrebi et al. (2018) combined predictive and corrective models, and
tested the feasibility and benefit of putting control strategies into practice. Instead of bunching
prediction, Arriagada et al (2019) used bus GPS data and smartcard data to investigate the causes of
bus bunching, with an emphasis on the planning side. Scheduled frequency, stop location and
configuration (number of the berths), traffic signal and bus lane design are found influential. This
research provides insight into bunching prevention in the planning stage.

3. Methodology

3.1 The identification of bus bunching event

As a bunching event involves two buses we refer to these as front bus and back bus respectively. Let
𝑛
a binary variable 𝑏𝑚 denote whether bus run m is caught in bunching as the back bus during its
𝑛 𝑛
dwelling at stop n. 𝑎𝑚 and 𝑑𝑚 denote the arrival and departure time of bus run m at stop n
respectively. At stop n, for each bus run 𝑚 (m≥2) we can obtain ∆𝑛𝑚−1,𝑚 which is the time interval
between the arrival time of bus m and the departure time of bus m-1 in Eq. (1). Bus run m is considered
bunched with bus run m-1 at the stop when ∆𝑛𝑚−1,𝑚 is below a threshold ∆0. The threshold can be
determined by the operator. Yu et al. (2016) and Moreira-Matias et al. (2016) used 1/4 of the scheduled
headway. ∆𝑛𝑚−1,𝑚 is defined as the departure-to-arrival headway in this study. Different from arrival-
to-arrival or departure-to-departure headway, ∆𝑛𝑚−1,𝑚 is negative when two buses overlap at the stop.
As overtaking is not allowed, for each stop n, bus m-1 always arrives and departs earlier than bus m,
and accordingly time interval ∆𝑛𝑚−1,𝑚 can always be obtained before the departure of bus m.

∆𝑛𝑚−1,𝑚 = 𝑎𝑚
𝑛 𝑛
− 𝑑𝑚−1 (1)

𝑛
For each bus m (m≥2), the binary bunching status 𝑏𝑚 can be derived by Eq. (2)

𝑛
1, ∆𝑛𝑚−1,𝑚 ≤ ∆0
𝑏𝑚 ={ (2)
0, ∆𝑛𝑚−1,𝑚 > ∆0

3.2 Variable selection

Following afore reviewed literature, the continuous ∆𝑛𝑚−1,𝑚 can be used as the dependent variable
for headway-prediction approaches. For bunching prediction then an additional step is required
judging whether the predicted headway is below a prior defined bunching threshold or not. Instead, in
𝑛
this study, 𝑏𝑚 is used as the dependent variable of the logistic regression to directly predict the binary
bunching status and bunching probabilities.
Gradually accumulated or suddenly significant inequality in dwell time and travel time might lead
two successive buses to be bunched. The back bus in a bunching event tends to have a shorter forward-
looking headway, negative deviation from timetable (ahead of schedule), less on-board passengers and
shorter dwell time than those of front buses in a bunching event or of non-bunched buses (Degeler
Heydenrijk-Ottens, Luo, Oort, & Lint, 2018). Yu et al. (2016) used boarding and alighting numbers of
two successive buses, link travel time and headway at an upstream stop as the input to their headway-
based prediction approach. As only bus GPS data is used in this study, information regarding boarding,
alighting as well as on-board passengers are not available. Instead dwell time is included in the variable
in addition to headway. Deviation from the timetable is excluded here, as bus dispatching is not based
on the timetable in some cities and the data for this variable might not be available. To conclude, dwell
time of two successive buses and their headway at an upstream stop n-k are used as the main leading
indicators of a coming bunching event in the k-step-ahead prediction. The detailed notation is as
follows:

prediction horizon in terms of number of stops, k = 1,2,3,…N-1 and N

k
denote the last stop of the bus route

𝑛−𝑘
𝑡𝑚 dwell time of bus run m at stop n-k

𝑛−𝑘
𝑡𝑚−1 dwell time of bus run m-1 at stop n-k

time interval between the arrival time of bus m and the departure time of
∆𝑛−𝑘
𝑚−1,𝑚
bus m-1 at stop n-k

We always have n>k, so that the k-stop-ahead prediction cannot be carried out until bus run m passes
the initial k bus stops, e.g. the prediction starts from stop 6 in the 5-stop-ahead prediction by using the
data at stop 1. Also note that m≥2 and that the first bus has zero probability to be bunched as the back
bus.

3.3 Logistic regression

Logistic regression (LOGR) modeling is widely used in classification problems. In binary
classification, it not only helps to categorize observations into positive or negative class but also
interprets the causality by producing the significance of each independent variable. Moreover, it
computes the probability of each observation to be in the positive or negative class. The binary
𝑛
bunching status from the perspective of the back bus 𝑏𝑚 (m≥2) is taken as the dependent variable.
𝑛−𝑘
𝑛−𝑘
𝑡𝑚 , 𝑡𝑚−1 , and ∆𝑛−𝑘 𝑛 𝑛−𝑘 𝑛−𝑘 𝑛−𝑘
𝑚−1,𝑚 are the independent variables. Let 𝑿𝑚 = [𝑡𝑚 , 𝑡𝑚−1 , ∆𝑚−1,𝑚 ], then the

probability of bus run m being bunched at stop n as a back bus can be derived as

𝑛 𝑛)
1
𝑃𝑟(𝑏𝑚 = 1|𝑋𝑚 = 𝑛 (3)
1 + 𝑒 −𝜷𝑿𝑚

𝑛
With parameters 𝜷 = [𝛽0 , 𝛽1 , 𝛽2 , 𝛽3 ] estimated by fitting the model with real data, 𝑃𝑟(𝑏𝑚 = 1|𝑿𝑛𝑚 )
for each bus run m (m≥2) at any stop n (n>k) can be computed k-stop ahead in the prediction stage.
𝑛 𝑛
𝑏𝑚 is predicted to be positive (one-event) if 𝑃𝑟(𝑏𝑚 = 1|𝑿𝑛𝑚 ) exceeds a probability threshold 𝑃𝑟𝑥
which is also known as the cut-off point, otherwise, negative (zero-event), as in Eq. (4).

𝑛
𝑛 1, 𝑃𝑟(𝑏𝑚 = 1|𝑿𝑛𝑚 ) > 𝑃𝑟𝑥
𝑏𝑚 ={ (4)
0, 𝑃𝑟(𝑏𝑚 = 1|𝑿𝑛𝑚 ) ≤ 𝑃𝑟𝑥
𝑛

3.4 Rare events bias

Irregular arrivals are common in bus transit operation, however few of them turn into severe bunching.
The prior bunching probability which is the ratio of bunching occurrence to the total number of
dwelling in Yu et al. (2016) varies from 3% to 17%, from 0.15% to 7.17% in Moreira-Matias et al.
(2016), and from 3% to 9% in our 5-day testing data. Bunching is hence a “rare” event in the dataset.
“Rare events data” refer to large datasets in which it is significantly less likely that the binary
dependent variables take one than zero. King and Zeng (2001) considered events such as wars, natural
disasters or epidemiological infections within long term time series data. They found logistic
regression underestimates the probability of rare events because they tend to be biased towards the
majority class, which is the less important class in most cases. This can be explained as follows:

The dependent variable 𝑌𝑖 follows a Bernoulli probability distribution that can take the values of
one and zero with probabilities 𝜋𝑖 and 1 − 𝜋𝑖 respectively. The probability function can be written
as

𝑃𝑟(𝑌𝑖 |𝜋𝑖 ) = 𝜋𝑖 𝑌𝑖 (1 − 𝜋𝑖 )1−𝑌𝑖 (5)

It is easy to derive the expectation and variance of 𝑌𝑖 as

𝐸(𝑌𝑖 ) = 𝜋𝑖 (6)
 𝑉(𝑌𝑖 ) = 𝜋𝑖 (1 − 𝜋𝑖 ) (7)

If the regression model has some explanatory power, the variance in the dependent variable has to
be large enough. The variance becomes larger as 𝜋𝑖 increases and reaches its maximum if 𝜋𝑖 = 0.5,
which indicates that it is favorable to involve an equal number of ones and zeros in the dataset. Cosslett
(1981) and Imbens (1992) also showed that equally sampling the two classes is optimal.

King and Zeng (2001) further discuss that selective data collection strategies instead of sampling
all available events could save data collection costs and correct the bias. Maalouf and Trafalis (2011)
implemented kernel logistic regression to rare events data, making use of a fast and robust adaptation
of kernel logistic regression and taking the weight of rare events into account. In this paper, selective
sampling and corresponding prior correction are used to reduce bias induced by rare events.

3.4.1 Sampling
Since bus GPS records are plentiful and easy to filter, efficient sampling thus can be achieved by
creating a balanced dataset in which all bunching events are included and part of the non-bunching
events are excluded. A balanced selection to include ones (bunching) and an equal number of zeros
(non-bunching) is applied.

3.4.2 Prior correction

Following King and Zeng (2001), prior correction is to correct the estimates according to the fraction
of ones in the population, denoted by τ, and the observed fraction of ones in the sample, denoted by
𝑦̅, since the probability of events to be predicted as ones is overestimated in the sample. The correction
is applied to the intercept 𝛽0 as

1−𝜏 𝑦̅
̂0 = 𝛽0 − ln [(
𝛽 )( )] (8)
𝜏 1 − 𝑦̅

3.5 LR and SVM as benchmark solutions

We now turn to two headway prediction methods that we consider as benchmarks compared to the
afore introduced direct bunching prediction method. Firstly, we consider linear regression (LR)
which is a basic tool in addressing prediction problems. To make LR comparable with LOGR, the
𝑛−𝑘
same set of independent variables 𝑿𝑛𝑚 = [𝑡𝑚
𝑛−𝑘
, 𝑡𝑚−1 , ∆𝑛−𝑘
𝑚−1,𝑚 ] is applied. With 𝜷′ =
[𝛽0′ , 𝛽1′ , 𝛽2′ , 𝛽3′ ] the relationship between the headway at stop n and the set of the independent variables
containing information k-stop-ahead is modeled as
 ∆𝑛𝑚−1,𝑚 = 𝜷′ 𝑿𝑛𝑚 (9)

Secondly, support vector machine (SVM) can map a non-linear relationship for model input and
output, and is tested by a number of studies in predicting bus headway or arrival time (B., Yu et al.,
2006; 2011; H., Yu et al., 2016). The same independent variables and dependent variable are applied
to the SVM regression, and a RBF (Radial Basis Function) kernel is selected because it is found both
efficient for bus arrival time prediction (Yu et al., 2011) and for bus headway prediction (Yu et al.,
2016).

4. Data description and case study settings

Buses are the main mode of public transport in Kyoto, Japan with more than 100 lines being served
by several operators. Bus GPS data of two primary bus operators has been obtained for a period of six
months in 2016. The data is collected every 8 seconds and provides the geographic coordinates of bus
location in real-time as well as associated bus line and vehicle number. Due to the lack of stop-based
information, it is essential to identify arrival and departure times for each bus run at each stop. Using
bus stop coordinates the distances of a bus from previous and next stops can be computed for every
GPS record. Considering that bunching and traffic congestion might make it difficult for the bus driver
to stop the bus at the exact bus stop coordinates as well as inaccuracy of GPS records, the bus is
regarded arriving at the stop once it approaches the bus stop within 30m. In the same way the departure
time is obtained when the GPS records indicate that the bus has moved 30m from the bus stop.
The data collection period includes the months of April and November. During these months, Kyoto
City experiences vast numbers of domestic and foreign visitors who come to enjoy the cherry blossoms
(April) and red leaves (November) in various sites around the city. The bus operators thus encounter
a huge challenge during these seasons to deliver a reliable service.
A circular bus line, Kyoto City Bus No. 205, which connects the city center, railway station and
several famous tourist attractions (Figure 1(middle)) is selected for the case study. There are 53 stops
on this bus line in total. To exclude the effect of dispatching at the terminal and factors for which we
do not have data (e.g. crew shedule, departure time adjustments), the 2nd stop of the line is taken as
the initial stop and the 52nd stop as the last one so that each bus run passes 51 bus stops. Data of five
weekdays in April 2016 are used as the training dataset and data of another five weekdays in the same
month are used for testing the model.
The scheduled headway varies from hour to hour, and the mean scheduled headway at the initial
stop is 6.97min from 6 am to 8 pm. The shortest scheduled headway is 3min at 7 am. Based on this,
1min is used for the bunching threshold as larger threshold can include headway variance that does
not lead to bunching.
 Adequate time is required to project a successful correction, in particular, if the control strategy is
based on manual communication between the dispatcher and the bus drivers. In this study, the
proposed approach is tested under a long prediction horizon of 10 stops or more which gives the
operator more than 15min to react since the mean stop-to-stop travel time is 1.77min.
Figure 2 illustrates the bus runs departing from the initial stop between 8 am and 10 am. Bunching
occurs frequently along the bus line. Bus runs that are involved in bunching as the back bus of two or
more buses at least once are denoted in red, and the front buses of a bunching sequence are denoted in
blue. Buses in green are not involved in any bunching. The headway fluctuation patterns of seven red
trajectories are demonstrated in Figure 3. Because of the bunching effect, the forward-looking
headway of back buses fluctuate within a small range, but always below one minute, once bunching
has been occurring giving further support to our threshold choice of one minute.

Figure 1. Data collected (left), data of Kyoto City Bus No. 205 (middle) and its configuration on
real map (right).

Figure 2. Trajectories of Kyoto City Bus No. 205 in one day of April 2016
Figure 3. Headway fluctuation along the line for bunched buses

5 Headway prediction

In the following case study, the headway prediction results derived by LR and SVM are discussed at
first including a comparison of these results. In Section 6 then the focus is on the bunching prediction
using these two methods as well as the newly proposed LOGR model. In the third part of our case
study we compare the measures using ROC curves.

5.1 Linear regression

Table 1 shows the estimation results of the fitted LR model. For all the prediction horizons, the
headway between the target bus and its front bus ∆𝑛−𝑘 𝑛−𝑘
𝑚−1,𝑚 and the dwell time of the target bus 𝑡𝑚

are always significant at 0.1% level and have positive signs. For short-term prediction the coefficient
of ∆𝑛−𝑘
𝑚−1,𝑚 is close to 1 and it begins to deviate from 1 as the prediction horizon increases. Meanwhile
𝑛−𝑘 𝑛−𝑘
the coefficient of 𝑡𝑚 increases gradually as the prediction horizon extends. 𝑡𝑚−1 is insignificant
in some cases, but it is still considered an important variable indicating at-stop activities and passenger
𝑛−𝑘
loads. Long 𝑡𝑚−1 may shorten the headway, but it sometimes results from in-vehicle crowding as
well as high boarding demand which may cause boarding failures that lead the following bus to dwell
𝑛−𝑘
longer and increase the headway, thus the sign of 𝑡𝑚−1 is inconclusive but mostly negative.
Table 1. Coefficients of the independent variables in the LR model
Prediction horizon Intercept ∆𝑛−𝑘
𝑚−1,𝑚
𝑛−𝑘
𝑡𝑚 𝑛−𝑘
𝑡𝑚−1 Adjusted R2
1-stop-ahead -0.3604*** 1.0009*** 0.7084*** 0.1247*** 0.9681
2 -0.3689*** 1.0026*** 0.7735*** 0.0381* 0.9431
3 -0.4890*** 1.0037*** 0.9786*** 0.0845*** 0.9173
4 -0.4664*** 1.0051*** 0.9611*** 0.0203 0.8922
5 -0.5081*** 1.0065*** 1.0315*** 0.0115 0.8673
6 -0.4676*** 1.0075*** 0.9974*** -0.0751* 0.8424
7 -0.5320*** 1.0092*** 1.0664*** -0.0247 0.8183
8 -0.5070*** 1.0108*** 1.0177*** -0.0738* 0.7942
9 -0.5458*** 1.0111*** 1.1231*** -0.0931* 0.7710
10 -0.5613*** 1.0123*** 1.1489*** -0.1072** 0.7474
11 -0.5937*** 1.0136*** 1.1834*** -0.0932* 0.7245
12 -0.6405*** 1.0160*** 1.2057*** -0.0459 0.7012
13 -0.6283*** 1.0186*** 1.1944*** -0.0935 0.6790
14 -0.6675*** 1.0212*** 1.2323*** -0.0886 0.6564
15 -0.6325*** 1.0235*** 1.2236*** -0.1681*** 0.6349
*** <= 0.001, ** <= 0.01, * <= 0.05

5.2 Support vector machine

The RBF function has two tuning parameters (C, γ) to enhance the predicting power of the SVM model.
C is the cost parameter to penalize the misclassifying of a sample. C thus controls the complexity of
the classifier; a high C may greatly bend the “prediction hyperplane” to avoid any misclassifying
(Cherkassky and Ma, 2004). γ is the inverse of the radius of influence by the samples selected as the
support vectors of the model. γ determines the influence of a single sample, a high γ thus may
reduce the radius and limit the generalization performance of the model. According to the findings on
bus arrival time prediction in Yu et al. (2011), 𝐶 ϵ [2-5, 25], γ ϵ [0.1, 0.3] are recommended for the
two parameters. In this paper, (22, 1) is set for the two parameters after a grid search in which γ = 1
performs better in our dataset.

5.3 Performance evaluation index

MAPE (Mean Absolute Percentage Errors) and RMSE (Root Mean Square Errors) are commonly used
to evaluate the prediction performance regarding exact value arrival time or headway prediction. Let
M and N denote the total number of bus runs and stops for a bus line, ∆𝑛−𝑘 ̂ 𝑛−𝑘
𝑚−1,𝑚 and ∆𝑚−1,𝑚 denote

the actual value and predicted value for headway, MAPE and RMSE are obtained respectively in Eq.
(10) and Eq. (11). In order to prevent the denominator being close to zero, we follow the method of
Yu et al. (2016) to calculate MAPE and use the mean of actual headways ∆̅ instead of ∆𝑛−𝑘
𝑚−1,𝑚

𝑁 𝑀
1 ∆𝑛−𝑘 ̂ 𝑛−𝑘
𝑚−1,𝑚 − ∆𝑚−1,𝑚
𝑀𝐴𝑃𝐸 = ∑ ∑| | × 100% (10)
(𝑀 − 1)(𝑁 − 𝑘) ∆̅
𝑛=𝑘+1 𝑚=2

𝑁 𝑀
1
𝑅𝑀𝑆𝐸 = √ ∑ ∑ (∆𝑛−𝑘 ̂ 𝑛−𝑘 2
𝑚−1,𝑚 − ∆𝑚−1,𝑚 )
(11)
(𝑀 − 1)(𝑁 − 𝑘)
𝑛=𝑘+1 𝑚=2

5.4 Performance comparison

Headway prediction results at Stop 23 “Kinkaku Temple”, one of the most frequented sightseeing
spots in Kyoto, is used to illustrate the performance of the aforementioned two methods. The results
of 1-stop-ahead and 10-stop-ahead predictions are illustrated in Figure 4 and evaluated in Figure 5.
Reliable prediction results (MAPE = 7.42% and RMSE = 0.71min by LR, MAPE = 7.45% and
RMSE = 0.71min by SVM) are produced for 1-stop-ahead prediction. For 10-stop-ahead prediction,
the results obviously deteriorate (MAPE = 21.64% and RMSE = 1.93min by LR, MAPE = 21.51%
and RMSE = 1.92min by SVM). We suggest they can still provide insights into expected fluctuation
patterns downstream, but the exact value is not reliable. Furthermore, neither in 1- nor 10-stop-ahead
prediction can these two methods perform favorably under the circumstance that the actual headway
becomes extremely short and bunching is going to happen, as is highlighted by the blue box in Figure
4. Furthermore, Figure 5 illustrates that in terms of MAPE and RMSE, both methods produce close
prediction accuracy and deteriorate similarly. Instead of significant increases in prediction errors,
evaluation metrics deteriorate gradually as the prediction horizon extends.
 (a) 1-stop-ahead headway prediction

(b) 10-stop-ahead headway prediction

Figure 4. Performance comparison in terms of exact headway value
 (a) Deterioration in MAPE as the prediction horizon extends

(b) Deterioration in RMSE as the prediction horizon extends

Figure 5. Performance comparison in terms of RMSE and MAPE under various prediction horizons

6 Bunching prediction

6.1 Logistic regression

We now focus on bunching prediction, firstly with logistic regression. Estimation results with and
without rare events bias correction are shown in Table 2. Adjusted McFadden’s R2 obtained by Eq.
(12) is selected to measure the overall goodness of fit for the logistic regression model.

2
𝑙𝑛𝐿𝑓𝑢𝑙𝑙 − 𝐾
𝑅𝑀𝐶𝐹 =1− (12)
𝑙𝑛𝐿𝑛𝑢𝑙𝑙
where 𝐿𝑓𝑢𝑙𝑙 is the likelihood derived by the fitted model, and 𝐿𝑛𝑢𝑙𝑙 is the likelihood of a null model
with intercept as the only predictor. K is the number of independent variables in the proposed model.
Due to the randomness generated by drawing non-bunching observations from the 5-day dataset to
correct the rare event bias, we run the model for 100 times and report the mean values for the
coefficients and adjusted McFadden’s R2. The significance is not based on any specific run but on all
the 100 runs, and for each variable the p-value is obtained by one sample t-test on the 100 estimated
coefficients. Bunching probability is negatively correlated with the value of headway, thus the
coefficients of the variables in the fitted LOGR have a reversed sign compared to those in the LR
2
model. The correction is proven effective as the adjusted 𝑅𝑀𝐶𝐹 is increased by at least 0.05 for each
prediction horizon.
Table 2. Coefficients of the independent variables in the LOGR model
Prediction horizon Intercept ∆𝑛−𝑘
𝑚−1,𝑚
𝑛−𝑘
𝑡𝑚 𝑛−𝑘
𝑡𝑚−1 2
Adjusted 𝑅𝑀𝐶𝐹
Without correction
1-stop-ahead 3.0842*** -2.2341*** -1.4015*** -0.3552*** 0.7508
2 2.3653*** -1.7302*** -1.0123*** -0.1913* 0.6899
3 2.0826*** -1.4265*** -1.1577*** -0.1057 0.6357
4 1.8370*** -1.2634*** -1.0737*** 0.0149 0.6016
5 1.6296*** -1.1234*** -0.9013*** -0.0036 0.5651
6 1.5044*** -1.0363*** -0.9266*** 0.1282 0.5385
7 1.4301*** -0.9596*** -0.8913*** 0.1007 0.5097
8 1.2780*** -0.8957*** -0.7406*** 0.1680* 0.4838
9 1.2385*** -0.8352*** -0.8607*** 0.1960** 0.4558
10 1.1162*** -0.7811*** -0.8076*** 0.2664*** 0.4289
11 1.1189*** -0.7361*** -0.7486*** 0.0807 0.4028
12 1.0504*** -0.7003*** -0.7425*** 0.1446* 0.3819
13 0.9763*** -0.6679*** -0.6612*** 0.1593* 0.3615
14 0.9184*** -0.6403*** -0.5862*** 0.1666** 0.3431
15 0.8498*** -0.6137*** -0.5155*** 0.1807** 0.3246

With correction (mean values of 100 runs reported with significance also based on all runs)
1-stop-ahead 2.3647*** -2.0343*** -1.1165*** 0.0619*** 0.8214
2 1.8730*** -1.5919*** -0.8407*** 0.0385* 0.7732
3 1.7571*** -1.3429*** -1.1465*** 0.0985*** 0.7267
4 1.6035*** -1.2226*** -0.9988*** 0.1627*** 0.6971
5 1.4231*** -1.0874*** -0.9060*** 0.1598*** 0.6568
6 1.2997*** -1.0042*** -0.8988*** 0.2585*** 0.6248
7 1.2399*** -0.9326*** -0.7756*** 0.1574*** 0.5938
8 1.1046*** -0.8738*** -0.7128*** 0.2740*** 0.5641
9 1.1111*** -0.8146*** -0.8855*** 0.2514*** 0.5311
10 1.0183*** -0.7663*** -0.8584*** 0.3289*** 0.5023
11 0.9848*** -0.7188*** -0.7544*** 0.1642*** 0.4709
12 0.9108*** -0.6817*** -0.7792*** 0.2616*** 0.4461
13 0.8268*** -0.6467*** -0.7202*** 0.2776*** 0.4195
14 0.8020*** -0.6183*** -0.6655*** 0.2246*** 0.3970
15 0.6966*** -0.5887*** -0.5887*** 0.2788*** 0.3739
*** <= 0.001, ** <= 0.01, * <= 0.05
6.2 Performance evaluation index
We define an actual bunching as “observed positive” and a predicted bunching as “predicted positive”.
Similarly, for non-bunching we define “observed negative” and “predicted negative”. All the
prediction results can be cast into four categories as is shown in Table 3, e.g. it is a true positive if an
observed bunching is correctly labeled one in the prediction outcomes. Four indexes can be obtained
from Eqs. (13) to (16). A binary classifier with high true positive rate and high true negative rate is
desired. The former is commonly referred to as “sensitivity” and the latter as “specificity”. Sensitivity,
specificity and accuracy which is an index computed with Eq. (17) to indicate overall prediction
performance, are applied to evaluate the binary classification performance of the three algorithms.

Table 3. Four categories for binary classification results

Observed positive (OP) Observed negative (ON)
Predicted positive (PP) True positive (TP) False positive (FP)
Predicted negative (PN) False negative (FN) True negative (TN)

∑ TP
True positive rate (TPR, sensitivity, SES) = (13)
∑ OP

∑ FP
False positive rate (FPR) = (14)
∑ ON

∑ TN
True negative rate (TNR, specificity, SPC) = (15)
∑ ON

∑ FN
False negative rate (FNR) = (16)
∑ OP

∑ TP +∑ TN
Accuracy (ACC) = (17)
∑ OP+∑ ON

For headway-based methods, only one combination of sensitivity and specificity is derived, as
headway prediction produces an exact value for each headway, resulting in deterministic true positive
and negative outcomes. Instead, by using logistic regression different combinations are obtained
depending on the cut-off point applied to the predicted probability. The cut-off point is the threshold
to determine the predicted positive. The event is judged as positive if its predicted probability exceeds
the cut-off point. A high cut-off point tends to only identify events presenting convincingly high
probability as positives, and consequently, it thus might misclassify observed positives as negative.
Vice versa, a low cut-off point will lead to more false positives. Therefore the cut-off point choice
should depend on the operator’s attitude towards bunching. Two scenarios are assumed here to
represent operators with different weights to false negative errors (missing actual bunching). Moreira-
Matias et al (2016) employed a large weight of 10:1 for false negative compared to false positive for
aggressive control purposes. We consider more moderate weights of 1:1 and 3:1.

Scenario 1 (LOGR-N): the operator is bunching-neutral, and gives equal weight to false
positive and false negative.
Scenario 2 (LOGR-A): the operator is bunching-averse, and gives a 3:1 weight to false
negative over false positive predictions.

The cost function in Eq. (18) computes the total weighted errors given a cut-off point. For LOGR-
N, 𝑤𝐹𝑃 = 𝑤𝐹𝑁 = 1, and for LOGR-A, 𝑤𝐹𝑃 = 1, 𝑤𝐹𝑁 = 3. The cut-off point generating the lowest
cost is taken as the optimal one. Based on the scenario-specific predicted positives and negatives, the
combination of sensitivity and specificity is determined.

𝑐 = 𝑤𝐹𝑃 ∑ FP + 𝑤𝐹𝑁 ∑ FN (18)

6.3 Performance comparison

Considering that the results derived by LR and SVM are similar, the comparison here is among SVM
and two distinguished scenarios based on LOGR. As is presented in Figure 6(a), most bunching events
can be detected 1-stop in advance by all three methods, and LOGR-A produces several false positives
because it applies a more aggressive strategy to potential bunching events. However, LOGR-A
significantly outperforms in 10-stop-ahead prediction, as is illustrated in Figure 6(b). LOGR-A
captures a number of observed positives that are misclassified by SVM and LOGR-N although it
generates a few more false positives.
A further comparison among two headway-based approaches and two scenarios of logistic
regression is demonstrated in Figure 7. Sensitivity, specificity and accuracy for the four methods under
various prediction horizons are presented. LOGR-A shows remarkable robustness in terms of
sensitivity. On the contrary to the obvious deterioration of the other three methods, the sensitivity of
LOGR-A keeps above 65% under all the prediction horizons. Besides, it only slightly underperforms
the other three methods in terms of specificity, indicating an acceptable trade-off cost. Non-bunching
events overwhelm bunching events in the daily operation, and a slight underperformance in specificity
might introduce a large number of false positive. The exact numbers of true positives, false positives,
true negatives, false negatives derived in the 5-day testing data are listed in Table 4. LOGR-N always
generates the least total errors (highest accuracy). LOGR-A always correctly detects most bunching
events (highest sensitivity) at the cost of most total errors (lowest accuracy). The notable advantage of
LOGR over the other two methods is its trade-off functionality. It can achieve highest overall accuracy
and can outperform the other methods in terms of sensitivity, although it cannot realize both objectives
simultaneously.

(a) 1-stop-ahead bunching prediction

(b) 10-stop-ahead bunching prediction

Figure 6. Performance comparison in terms of binary bunching identification
 (a) Deterioration in SES as the prediction horizon extends

(b) Deterioration in SPC as the prediction horizon extends

(c) Deterioration in ACC as the prediction horizon extends

Figure 7. Performance comparison in terms of SES, SPC and ACC under various prediction
horizons
Table 4. Performance comparison for 10-stop-ahead bunching prediction
PP PN SES SPC ACC
Size OP
TP FP TN FN (%) (%) (%)
Day 1
LR 4223 344 161 77 3802 183 46.80 98.01 93.84
SVM 4223 344 171 89 3790 173 49.71 97.71 93.80
LOGR-N 4223 344 208 118 3761 136 60.47 96.96 93.99
LOGR-A 4223 344 262 200 3679 82 76.16 94.84 93.32

Day 2
LR 4182 384 158 118 3680 226 41.15 96.89 91.77
SVM 4182 384 166 138 3660 218 43.23 96.37 91.49
LOGR-N 4182 384 143 94 3704 241 37.24 97.53 91.99
LOGR-A 4182 384 266 345 3453 118 69.27 90.92 88.93

Day 3
LR 4264 355 175 62 3847 180 49.30 98.41 94.32
SVM 4264 355 187 73 3836 168 52.68 98.13 94.35
LOGR-N 4264 355 201 82 3827 154 56.62 97.90 94.47
LOGR-A 4264 355 263 202 3707 92 74.08 94.83 93.11

Day 4
LR 4182 146 51 42 3994 95 34.93 98.96 96.72
SVM 4182 146 57 51 3985 89 39.04 98.74 96.65
LOGR-N 4182 146 50 37 3999 96 34.25 99.08 96.82
LOGR-A 4182 146 100 112 3924 46 68.49 97.22 96.22

Day 5
LR 4182 254 123 67 3861 131 48.43 98.29 95.27
SVM 4182 254 121 79 3849 133 47.64 97.99 94.93
LOGR-N 4182 254 176 115 3813 78 69.29 97.07 95.38
LOGR-A 4182 254 204 164 3764 50 80.31 95.82 94.88
7 Discussion on the trade-off between sensitivity and specificity

ROC curves created by plotting (1-SPC, SES) for given cut-off points are commonly used to evaluate
the classification performance. AUC (Area Under the Curve) being close to one indicates good
classification power. ROC curves under various prediction horizons are presented in Figure 8.
Furthermore, the four combinations of sensitivity and specificity derived by the four methods
discussed in the previous section are indicated on each curve.
For each horizon, the corresponding curve can be considered the optimal front derived by LOGR.
If an algorithm outperforms LOGR, the point it represents should appear above the curve with a higher
SES and lower 1-SPC. It can be observed that the two headway-based methods (LR and SVM)
generally fall below and sometimes on the LOGR curve, although the downward deviation from the
curve is not significant.
It is easy to conduct the trade-off between sensitivity and specificity on the LOGR curve. The LOGR
curve contains all combinations of prediction performance given continuous cut-off points where each
cut-off point can be considered as optimal. A bunching-averse operator who is aggressive to eliminate
bunching might desire to detect 99% of the positives regardless of the cost to increase false positive
rate. This trade-off functionality significantly enhances the flexibility and robustness of existing
bunching prediction approaches, especially for putting the predictive methodology into real practice.
The curves provide a robust benchmark and insights for future algorithms that address bunching
prediction problem. Deterministic methods can only produce one combination of prediction
performance which greatly limits its contribution to the real application unless its sensitivity and
specificity simultaneously achieve a highly reliable level. Other probabilistic methods generating a
curve having higher AUC than LOGR or deterministic methods producing points of substantial
upward deviation from the curve under various prediction horizons should be further promising
extensions.
Figure 8. ROC curves under various prediction horizons (1-stop, 5-stop, 10-stop and 15-stop ahead)

Table 5. Supplementary information for ROC curves shown in Figure 8

AUC (area under the Cut-off point of Cut-off point of
Prediction horizon
curve) LOGR-N (%) LOGR-A (%)
1-stop-ahead 0.9922 90.31 73.50
5 0.9763 87.19 77.53
10 0.9546 87.56 79.74
15 0.9279 81.60 71.50
8. Conclusion and further work

In this study, the potential of logistic regression to predict bus bunching is discussed. We consider the
“rare event” nature of our problem which leads logistic regression to lose prediction power due to
being biased to the majority in the dataset where positive events are by far outnumbered by negative
events. Thus a selective sampling method and intercept correction is applied. We then compare this
method with existing approaches that predict headways and then utilize the headway prediction for
bunching prediction. Clearly headway prediction can be used for a larger range of purposes and deeper
understanding of the service regularity developments as well as control strategies. However, bunching
prediction in itself is important as it can be considered a distinctive state. This paper and other literature
illustrate that headways fluctuate, but that, once bunching is reached, this state mostly continues along
the line with far less headway fluctuation. We illustrate that when it comes to predicting bunching
itself the newly proposed method has the potential to outperform headway-based methods such as LR
and SVM in several aspects.
Firstly, LOGR provides superior prediction results under a long prediction horizon. It outperforms
LR and SVM by 28% in sensitivity and maintains the same level of specificity in 10-stop-ahead
prediction. It also shows improved resistance against deterioration in prediction performance as the
prediction horizon extends.
Secondly, robustness and flexibility are significantly enhanced. LOGR provides robust prediction
results that contain various sets of bunching outcomes under different cut-off points. This enables the
operator to apply weights that are in accordance with their attitude towards bunching and operation
budget. Some operators with limited possibility or willingness to apply corrective measures can use
SVM or LOGR with neutral cut-off point setting. On the contrary, operators who desire to eliminate
any possible bunching might be unwilling to choose headway-based methods which omit a
considerable number of bunching in the long-term prediction cases. In this case LOGR-A becomes a
much-preferred option. To conclude, LOGR provides operators with a wide range of options that can
be tailored by their attitudes towards unexpected system disturbances.

We find that the headway-predicting approaches deviate slightly downward from the optimal front
and we discuss that their shortcomings are inadequate robustness and flexibility from the operator’s
perspective. We note that it is also feasible to form a curve in terms of sensitivity and specificity for
probabilistic headway prediction methods with confidence intervals to realize the trade-off on the
curve discussed in this paper. Hans et al. (2015) developed a simulation-based prediction tool, Yu et
al. (2017) tested RVM algorithm on headway prediction problem. Both methods could be extended to
compute the probability of a headway falling below 1min and then to construct the ROC curves. By
doing so and comparing the different ROC curves more insights might be obtained.
 Finally, other extensions that potentially strengthen the predictive power of the models presented in
this study should be noted. The model itself has space for improvement by including variables such as
weather, traffic signals and passenger demand that are not incorporated due to missing data.
Furthermore, the study could be extended to simultaneously predict bunching for several lines, in
which case common line effects such as the interaction between buses of different lines at a common
stop need to be considered.

Acknowledgements

We acknowledge the contribution of Hiroshi Shimamoto who provided valuable insights in the early
stage of this research. We also acknowledge the comments from three anonymous reviewers that
helped us improve this paper substantially. Besides, we gratefully acknowledge ASTEM (Advanced
Science, Technology and Management Research Institute of Kyoto), notably Hideyuki Yamauchi, for
his support in providing us with the bus GPS data.

References

Andres, M., & Nair, R. (2017) A predictive-control framework to address bus bunching.
Transportation Research Part B, 104, 123-148.
Arriagada J., Gschwender, A., Munizaga, M., & Trepanier, M. (2019). Modeling bus bunching using
massive location and fare collection data. Journal of Intelligent Transportation Systems, 23(4),
332-344.
Bartholdi, J. J., & Eisenstein, D.D. (2012). A self-coordinating bus route to resist bus bunching.
Transportation Research, Part B, 46(4), 481-491.
Berrebi, S. J., Watkins, K. E., & Laval, J. A. (2015). A real-time bus dispatching policy to minimize
passenger wait on a high frequency route. Transportation Research Part B, 81, 377-389.
Berrebi, S. J., Hans, E., Chiabaut, N., Laval, J. A., Leclercq, L., & Watkins, K. E. (2018). Comparing
bus holding methods with and without real-time predictions. Transportation Research Part C, 87,
197-211.
Cherkassky, V., & Ma, Y. (2004). Practical selection of SVM parameters and noise estimation for SVM
regression. Neural networks 17(1), 113-126.
Cosslett, S. R. (1981). Maximum Likelihood Estimator for Choice-Based Samples. Econometrica,
49(5), 1289-1316.
Daganzo, C.F. (2009). A headway-based approach to eliminate bus bunching: Systematic analysis and
comparisons. Transportation Research Part B, 43(10), 913-921.
Daganzo, C.F., & Pilachowski, J. (2011). Reducing bunching with bus-to-bus cooperation.
Transportation Research Part B, 45(1), 267-277.
Dai, Z., Ma, X., & Chen, X. (2019). Bus travel time modelling using GPS probe and smart card data:
A probabilistic approach considering link travel time and station dwell time. Journal of Intelligent
Transportation Systems, 23(2), 175-190.
Degeler, V., Heydenrijk-Ottens, L., Luo, D., Oort, N., & Lint, H. (2018, July). Unsupervised approach
to public transport bunching swings formations phenomenon analysis. Paper presented at the
14th International Conference on Advanced Systems in Public Transport (CASPT), Brisbane,
Australia.
Eberlein, X.J., Wilson, M.H.M., & Bernstein, D. (2001). The holding problem with real-time
information available. Transportation Science, 35(1), 1-18.
Hans, E., Chiabaut, N., Leclercq, L., & Bertini, R. L. (2015). Real-time bus route state forecasting
using partical filter and mesoscopic modeling. Transportation Research Part C, 61, 121-140.
Imbens, G. (1992). An Efficient Method of Moments Estimator for Discrete Choice Models with
Choice-Based Sampling. Econometrica, 60(5), 1187-1214.
King, G., & Zeng, L. (2001). Logistic Regression in Rare Events Data. Political Analysis, 9(2), 137-
163.
Kumar, B. A., Vanajakshi, L., & Subramanian, S. C. (2018). A hybrid model based method for bus
travel time estimation. Journal of Intelligent Transportation Systems, 22(5), 390-406.
Maalouf, M., & Trafalis, T.B. (2011). Robust weighted kernel logistic regression in imbalanced and
rare events data. Computational Statistics & Data Analysis, 55(1), 168-183.
Moreira-Matias, L., Cats, O., Gama, J., Mendes-Moreira, J., & De Sousa, J.F. (2016). An online
learning approach to eliminate Bus Bunching in real-time. Applied Soft Computing, 47, 460-482.
Newell, G.F., & Potts, R.B. (1964). Maintaining a bus schedule. Proceedings of 2nd Australian Road
Research Board, 2, 388-393.
Newell, G.F. (1974). Control of pairing of vehicles on a public transportation route, two vehicles, one
control point. Transportation Science, 8(3), 248-264.
Osuna, E.E., & Newell, G.F. (1972). Control strategies for an idealized bus system. Transportation
Science, 6(1), 52-71.
Schmöcker, J. D., Sun, W., Fonzone, A., & Liu, R. (2016). Bus bunching along a corridor served by
two lines. Transportation Research Part B, 93, 300-317.
Sun, A., & Hickman, M. (2005). The real-time stop-skipping problem. Journal of Intelligent
Transportation Systems, 9(2), 91-109.
Sun, W., & Schmöcker, J. D. (2018). Considering passenger choices and overtaking in the bus
bunching problem. Transportmetrica B, 6(2), 151-168.
Wu, W., Liu, R., & Jin, W. (2017). Modelling bus bunching and holding control with vehicle
overtaking and distributed passenger boarding behaviour. Transportation Research Part B, 104,
175-197.
 Xuan, Y., Argote, J., & Daganzo, C.F. (2011). Dynamic bus holding strategies for schedule reliability:
Optimal linear control and performance analysis. Transportation Research Part B, 45(10), 1831-
1845.
Yu, B., Yang, Z., & Yao, B. (2006). Bus arrival time prediction using support vector machines. Journal
of Intelligent Transportation Systems, 10(4), 151-158.
Yu, B., Lam, W. H., & Tam, M. L. (2011). Bus arrival time prediction at bus stop with multiple routes.
Transportation Research Part C, 19(6), 1157-1170.
Yu, H., Chen, D., Wu, Z., Ma, X., & Wang, Y. (2016). Headway-based bus bunching prediction using
transit smart card data. Transportation Research Part C, 72, 45-59.
Yu, H., Wu, Z., Chen, D., & Ma, X. (2017). Probabilistic prediction of bus headway using relevance
vector machine regression. IEEE Transactions on Intelligent Transportation Systems, 18(7),
1772-1781.
Zhang, S., & Lo, H. (2018). Two-way-looking self-equalizing headway control for bus operations.
Transportation Research Part B, 110, 280-301.

View publication stats