Design and Development of an Application for Predicting Bus Travel

Times using a Segmentation Approach

Ankhit Pandurangi, Clare Byrne, Candis Anderson, Enxi Cui and Gavin McArdle

School of Computer Science, University College Dublin, Belﬁeld, Ireland

Keywords:

Travel Time Prediction, Bus Transport, Spatial Data, Machine Learning.

Abstract:

Public transportation applications today face a unique challenge: Providing easy-to-use and intuitive design,

while at the same time giving the end user the most updated and accurate information possible. Applications

often sacriﬁce one for the other, ﬁnding it hard to balance the two. Furthermore, accurately predicting travel

times for public transport is a non-trivial task. Taking factors such as trafﬁc, weather, or delays into account

is a complex challenge. This paper describes a data driven analysis approach to resolve this problem by using

machine learning to estimate the travel time of buses and places the results in a user-friendly application. In

particular, this paper discusses a predictive model which estimates the travel time between pairs of bus stops.

The approach is validated using data from the bus network in Dublin, Ireland. While the evaluation of the

predictive models show that journey segment predictions are less accurate than the prediction of a bus route in

full, the segmented approach gives the user more ﬂexibility in planning a journey.

1 INTRODUCTION

Due to its importance in trip and route planning, travel

time prediction has a long history and has been con-

sidered by several researchers in different domains

(Lin et al., 2005). There is no doubt regarding the

value of accurate prediction. While travel time pre-

diction is relevant to all transport modes, it is per-

haps road transport which receives the most atten-

tion. Studies have long shown that passengers are

demanding services which are informative and dy-

namic (Peek and van Hagen, 2002) and provide them

with the ability to factor waiting times and delays into

their trip planning (Kroes and Daly, 2018). From an

operators perspective, the ability to understand travel

times is important for scheduling and service pro-

vision (Gkiotsalitis and Cats, 2017). Despite much

activity in this area, many questions remain unan-

swered. This is due in part to the difﬁculties of in-

corporating the impact of external factors such as de-

mand, time of day, weather conditions or seasonal

patterns into prediction algorithms (Crist

obal et al.,

2019).

The aim of the work presented in this paper is

to provide an effective solution to the problem that

Dublin Bus, and other public transportation applica-

tions, face in relation to providing accurate estimates

of travel and arrival times for trip planning within an

easy to use application. Static journey predictions are

common. These provide a simple estimated travel

time for each route. This can be produced based on

a single average of all travel times. Such a time is

often unrealistic given the variables that impact travel

time such as demand, time of day, weather conditions

or seasonal patterns. In this paper we describe a new

algorithm for predicting travel times and integrate it

into an innovative mobile application. Predictions are

made by analysing historical weather and travel time

data with a Random Forest machine learning algo-

rithm based on estimating travel time between pairs

of adjacent bus stops. Random Forest Classiﬁcation

is an ensemble method in which a collection of indi-

vidual decision trees are used to produce a classiﬁ-

cation. The classiﬁcation which is produced my the

majority of trees in the ensemble is the output classi-

ﬁcation (Breiman, 2001). The approach has shown to

outperform individual constituent models.

The remainder of this paper is organised as fol-

lows: Section 2 examines related work in the area of

machine learning for predicting travel time as well as

a review of bus trip planning applications. Section 3

presents our approach for estimating bus travel time

between stops and also brieﬂy describes the interface

of the accompanying application. Section 4 presents

the results of evaluating the accuracy of the travel time

prediction. Finally Section 5 concludes the paper with

Pandurangi, A., Byrne, C., Anderson, C., Cui, E. and McArdle, G.

Design and Development of an Application for Predicting Bus Travel Times using a Segmentation Approach.

DOI: 10.5220/0009393800720080

In Proceedings of the 6th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2020), pages 72-80

ISBN: 978-989-758-425-1

a discussion and the identiﬁcation of some area for fu-

ture research.

2 RELATED WORK

Due to the importance of trip planning for both oper-

ators and end users there are many examples of trip

planning application across many domains including

commuting and tourism (Jariyasunant et al., 2011),

(Huang and Peng, 2002), (Brilhante et al., 2015). The

applications all have a user interface which allows the

location for origin and destination to be speciﬁed and

then provides a route recommendation using various

modes of transport considering different spatiotempo-

ral constraints. Some utilise real-time data (Liebig

et al., 2017) but the majority of applications use static

data and provide travel time estimate based on the best

or worst case. This section brieﬂy examines applica-

tions available before discussing techniques for more

robust travel time prediction using machine learning.

2.1 Bus Applications

There are a plethora of bus and transport planning

applications available to support travellers and com-

muters. A review of other transport apps currently

available on mobile app stores was carried out. Most

allow users to plan routes, view timetables, locate bus

stops and ﬁnd points of interest. In relation to jour-

ney time predictions, several apps integrated Google’s

journey prediction API, rather than building their own

model; examples include the Hit the Road, SDMTS &

NCTD, MuniMobile, CityMapper and OneBusAway

Seattle apps. It is unclear if any of the apps used any

more criteria in their predictions, other than the search

terms in the journey planner.

In the Irish context, the Dublin Bus App makes

extensive use of real-time information to estimate the

arrival time of a bus. The application also includes

journey planning, favourites, and fare calculating fea-

tures; but does not provide journey time predictions.

Instead, travel times are given by timetables which

give journey estimates for several sections of the jour-

ney (for example, the number 4 route is broken into

seven journey segments, with an estimate for each).

Static predictions like these do not reﬂect likely trafﬁc

conditions for a given journey and are not very use-

ful to the user. Transport for Ireland (TFI) have de-

veloped a Real-time Journey Planner application for

Dublin. When queried for a particular destination, the

application returns a multi-modal journey plan with a

travel time prediction for each mode segment in the

journey. These estimates are dynamic, and predic-

tions take into account factors such as time of day.

It is not clear if weather-related factors are taken into

account. This is a stand-alone app for journey plan-

ning and does not have a diverse range of features in

comparison with other apps analysed such as ticket

top up options.

2.2 Estimating Bus Travel Time

Recently, advances in machine learning along with

the availability of new data sets have renewed interest

in data driven approaches for predicting travel time

in transport systems. Several machine learning ap-

proaches have been investigated for this problem in-

cluding Support Vector Machine (SVM) and Artiﬁcial

Neural Network (ANN) models (Chang et al., 2010).

However, both approaches can take a signiﬁcant time

to train which may make them unsuitable for real-time

predictions(Lee et al., 2012). Mazloumi et al. (2011)

investigated these techniques for bus routes in Mel-

bourne, Australia and decided to prioritise ANN mod-

els due to their modelling ﬂexibility, predictive abil-

ity and generalization potential. Ensemble methods

such as Random Forest models were also explored by

several researchers (Gal et al., 2017). These models

are generally faster to train and can be optimised rel-

atively easily, and thus have potential to scale (Gal

et al., 2017). Hybrid models also appear in the lit-

erature, but these models either showed mediocre re-

sults or had a signiﬁcantly long processing time (Peng

et al., 2018),(Petersen et al., 2019). Linear Regres-

sion models were tested by Jeong and Rilett (2004)

and compared to Artiﬁcial Neural Network Models

and historical averages, with the result that they were

worse performing than both and therefore ruled out

of consideration. For the interested reader, Crist

obal

et al. (2019) provide an excellent review of these ap-

proaches applied to bus travel time prediction.

Within travel time prediction, there are different

approaches in terms of what time is to be predicted,

the whole route or segments. Within this context there

are also different approaches to deﬁning segments of

routes for analysis. Celan (2017) examined the under-

lying trafﬁc system to determine optimal segments.

This study compared a model “deﬁned by bus stops

and crossings of the road network”, a model “deﬁned

by bus stops”, “a data model which addresses the in-

dividual parts of the network in relation to the poten-

tial barriers that affect the travel speed of buses”, and

“a data model with ﬁxed-length links of the bus net-

work”. Many of these models required a signiﬁcant

amount of contextual information to deﬁne bus stop

segments which adds additional overhead to the work.

Gal et al. (2017) proposes a segmentation technique

Design and Development of an Application for Predicting Bus Travel Times using a Segmentation Approach

where a segment is deﬁned by pairs of bus stops. His-

torical data are used with an approach based on queu-

ing theory to predict journey time. Segments are com-

bined to produce predictions for larger trips. The ap-

proach was tested with data with the bus network in

Dublin. These segmentation techniques and the need

to understand the network inﬂuenced our approach

which is described in the next section.

While ANN predicted accurate results when com-

pared to more classical approaches such as linear re-

gression, there is a trade off when compared to the

computation time. Models built using Random For-

est approaches are faster to train and can handle large

feature sets while still producing comparable results

(Gal et al., 2017). In Section 3 the development of

the model is described which uses segments based on

pairs of adjacent bus stops. The evaluation metrics

used in the literature include root-mean squared error,

mean absolute relative error and median absolute rel-

ative error. These measures will be also used to assess

the effectiveness of our approach.

3 APPROACH

While the Dublin Bus App is widely used across

Dublin and is most likely the most used application

for checking bus arrival times, it is not built with ac-

curate times in mind. The use of static timetables

rather than a predictive model means that a user could

be waiting at a bus stop for 10 minutes, while the

app claims that a bus is simply 1 minute away. Our

approach to the problem relies on an understanding

of how Dublin’s bus system is designed. Unlike a

city like London, where buses have multiple hubs and

routes that are centred around various points (for ex-

ample, Victoria Station, Euston Station, etc.), Dublin

relies on the majority of bus routes passing through

the city centre. In this sense, it is more of a “star”

design rather than a hub design. Of course, the city

centre is the area most likely to be congested which

adds to the uncertainty in travel time prediction. How-

ever, this also means that most buses will pass through

the same set of stations at some point, meaning that

rather than build a model that tries to speciﬁcally tar-

get bus routes, our approach is to build a model that

uses stop-pairs to predict journey time. For example,

bus routes 39a, 46, and 145 all pass through the same

set of stops for a signiﬁcant portion of their respective

routes. Therefore, instead of developing three sepa-

rate models for predicting travel time, one model that

handles common element of their routes is proposed.

The predictive aspects of the approach are incorpo-

rated into a mobile web application which functions

for single bus or multi-connection routes, and jour-

neys that utilise the entire route or just a few stops of

a route. It is fully scalable and can handle the variety

of routes an average user might want to take.

3.1 System Description

The ﬁnal product is a transportation app which brings

together a journey planner with prediction, real-time

bus information, the ability to log in and save favorite

routes, a game, and route list. The map contains fea-

tures such as a user ﬁnding their nearest stops, and

planning a route to a selected location. There are ﬁve

sections of the app - Home, Journey Planner, Routes,

Favourites, and More.

3.1.1 Wire Frame Design

During the early planning stages, wire frames (see

Figure 1) were created to show how the app would

like and to aid the design. The ﬁnal product was quite

similar to these wire frames and is discussed in the

coming sections.

Figure 1: Original Wireframes.

3.1.2 Home Page

The home page, shown in Figure 2, is map-centric and

by default shows the user’s location and nearby stops.

3.1.3 Real-time Information

As shown in Figure 3, there are markers on the home

and journey planner pages which show where Dublin

GISTAM 2020 - 6th International Conference on Geographical Information Systems Theory, Applications and Management

Figure 2: Home Page.

bus stops are located. When a user clicks one of the

markers, an information window displays the stop ID,

stop name and the routes which stop at that location.

There is a real-time button, which when pressed, de-

tails bus arrival times at that location - taken from the

Real Time Passenger Information (RTPI) API, sup-

plied by Dublin Bus.

Figure 3: Markers and Real-time Information.

3.1.4 Journey Planner and Search Pages

The journey planner is map-centric and allows for

search and planning of journeys. Once a user types

in an address or business of their choice and hits

“search”, pin(s) will appear on the map. These mark-

ers change based on the type of business for which

they search (restaurant, hotel, school, etc), as pre-

sented in Figure 4. Clicking that icon or marker and

then “route to here” presents the user with the pos-

sible bus routes to that place from the user’s current

location.

3.1.5 Planning a Journey

The “show journey planner” button allows the user to

input an origin and destination, along with a date and

time for travel. Once submitted, it displays different

possible routes - with the number of connections re-

quired – along with an estimated time for the journey

Figure 4: Search and Journey Planning.

(described in Section 3.2. When one of the routes is

clicked, the map then displays each stop on the jour-

ney along with detailed directions including any con-

nections and estimated walking time. This is demon-

strated in Figure 5 for both the desktop and mobile

versions.

Figure 5: Desktop and Mobile Journey Planner and Predic-

tion.

Design and Development of an Application for Predicting Bus Travel Times using a Segmentation Approach

3.1.6 Routes

The app contains a page which lists all Dublin Bus

routes with numbers and destinations. For ease of

use, the user may search for their required route by

its number. Figure 6 demonstrates that when a user

clicks on a particular route, a map appears with all of

the route’s stops listed, in both directions.

Figure 6: Route Information.

3.1.7 Favourites

As seen in Figure 7, there is login functionality built

into the favourites page which allows the user to

access favourites (as well as the game) only when

logged in. The user is required to provide basic in-

formation - username, email, and password - to sign

up. On the favourites page, the user is able to cre-

ate and delete favourite journeys, as well as plan that

particular trip.

3.1.8 Additional Functionality

As part of its innovation, the application contains a

game (seen in Figure 8) in which twenty Dublin land-

marks are marked on a map. The user can click the

marker to learn more about the attraction - and if

the (logged-in) user visits the attraction they receive

points. These points are displayed on the more page,

along with cumulative calculations of CO

savings,

a weather forecast, and a news feed from the Dublin

Bus website.

Figure 7: Signing Up and Favourites.

Figure 8: Game Marker and More Page.

3.2 Data Management and Feature

Selection

Dublin Bus provided 12 months of historical infor-

mation on vehicles, individual trips, and GPS data

for each stop. Weather information was also gath-

ered from data.gov.ie, which was cross-referenced

with the historical bus data. The weather data con-

tained the weather conditions every hour for a mul-

titude of different categories. This included features

like the amount of rain (mm), the temperature, and

the wind-speed. These were the main three features

of the weather information that we felt would impact

the general population’s use of the public transport

system, and were used for our machine learning mod-

els. The bus data was more complicated. It was split

across two ﬁles, each containing a part of the data re-

lating to a speciﬁc trip made on a route, the duration

GISTAM 2020 - 6th International Conference on Geographical Information Systems Theory, Applications and Management

of the route, the planned departure and arrival times,

and the actual departure and arrival times. There were

also columns that contained ID’s for various trip parts,

stops, etc. The ﬁnal dataset used in the machine learn-

ing contained 50 columns, a mix of data from the Na-

tion Transport Authority, weather data, and columns

created by the team to signify certain things, such as

whether the trip was during peak times, whether it was

heavy rain or light rain, etc.

As the data provided by Dublin Bus was large,

working with it was a particular challenge. The Pan-

das suite of data analysis tools

was used in conjunc-

tion with early ﬁltering and sampling in order to man-

age the data. To clean the trip data, columns with un-

usable data were dropped, and null and contradictory

values were investigated and dropped. It was neces-

sary to drop rows that had been suppressed so that

they wouldn’t skew the data set. Many features for

the ‘Trips’ data were built and included day of week

(Monday-Sunday), if the day was a holiday, as well as

rush hour. For rush hour, the actual departures from

7:00-8:30am and 4:00-6:00pm were used because if

trips average between 30 minutes and 1 hour, all of

the “rush hour” trips would be included. Columns

were added for the journey duration as well as the

difference in duration between the actual and planned

journeys. To merge the ‘Trips’ data with the weather

data, it was crucial to create a feature to combine

the date with the actual time of arrival - so that the

hourly weather could be matched with the hour the

bus arrived. Dummy variables were created for the

rain measurements, including current and past rain.

These categories were then binned by type (ex: rain,

snow, shower, clouds, etc). Figure 9 shows the origi-

nal weather columns.

Figure 9: The Weather Columns in the Weather Dataset.

The analysis showed there was very little correlation

on any of the features when plotted against the ‘dura-

tion difference’, as you can see in Figure 10. In order

to improve this, additional features on the ‘Trips’ data

frame were created for weekday rush hour, Friday,

https://pandas.pydata.org/

Saturday, and Sunday. New weather features were

also created - precipitation now (combining together

rain, snow, drizzle, etc), precipitation in the past hour,

and a feature that combined precipitation now with

precipitation in the past hour.

Figure 10: Correlation with the ‘Duration Difference’.

Each of the features was plotted against the ‘dura-

tion difference’ value. No single feature was iden-

tiﬁed as a good predictor; but the plots showed that

while there are some outliers, there are not many,

which is useful information when cleaning the data.

As the correlation between features and the ‘duration

difference’ was so poor, pairs of features were ex-

amined as shown in Figure 11. Again, this did not

lead to a greater insight. Instead of using any of the

pairs for modeling, features with the highest correla-

tion were chosen - the ﬁnal features used were ‘day-

OfWeek’(with 0-6 being Monday through Sunday),

‘friday’ (boolean), ‘rushHour’ (boolean), ‘temp’ (in

degrees C), ‘monToThurRushHour’ (boolean), ‘direc-

tion’ (either 1 or 2), and ‘windSpeed’.

Design and Development of an Application for Predicting Bus Travel Times using a Segmentation Approach

Figure 11: Plots Against ‘Duration Difference’.

3.3 Prediction Algorithm

Two approaches to predicting travel time were imple-

mented. An early approach taken was to predict the

travel time for a full route. Tests showed that using

a Random Forest classiﬁcation with the features de-

scribed above to predict travel time produced a model

with a Root Mean Square Error (RMSE) of 600 sec-

onds, or 10 seconds per stop. This was not a satis-

factory outcome so a second approach was employed.

This strategy involved taking a percentage of that es-

timated time to present to the user for speciﬁed stop

segments, but accuracy was severely limited. The

next steps were to decrease the RMSE and Mean Ab-

solute Error, increase the R, and create models efﬁ-

ciently.

A study by Jiang (2017) that worked with sim-

ilar historical bus data concluded that “whole route

travel time prediction using segments has better re-

sults than the route prediction model using solely bus

GPS data”. The potential increase in accuracy paired

with the ability to predict segments of trips led to the

use of a segmented approach in model creation. This

segmentation strategy allowed the data to be turned

into a ‘network’ of linked stops, rather than a list of

individual stops. This ‘network’ approach meant that

full route information for each bus line could be gen-

erated easily, including each stop id, its ‘progrnum-

ber’ in the route, and the latitude and longitude of

each stop. This information then formed the basis for

the app’s routes page.

In order to calculate journey segments, as well

as the full journey duration time, each journey was

divided into ‘stop pairs’. Each row would need to

record the time taken to travel from the ﬁrst to the

second stop in the pair, for each stop pair in a speciﬁc

trip. The ‘progrnumber’ feature allowed the ability to

arrange stop pairs in the correct order of the trip. As

processing all of the leave times data (over 116 mil-

lion rows) would have taken several weeks to process,

a sample of 12,000 trip IDs was taken, divided evenly

across each direction. The goal was to create a new

collection of ‘stop pairs’ and their travel time dura-

tion, rather than the individual ‘stop’ information that

already existed. This enabled the app to return predic-

tions for a segment of a route by adding all stop pair

predictions between a speciﬁc start point and destina-

tion. It also increased the amount of data available,

as many stops exist on several routes, and this extra

data had the potential to increase the accuracy of the

models.

Several routes were not included in the sampling,

or had limited amounts of data, and required a more

hands-on approach. For roughly twenty routes, addi-

tional individual ﬁles were created. For a single route,

for one direction, the data set held between 3,000 -

7,000 rows of data; however, as previously stated,

some data sets were smaller. To increase the accu-

racy of the models, two different model types were

used - Random Forest and linear regression, both im-

plemented using the scikit-learn package

. Complex

model types are generally not suited to dealing with

such small amounts of data, thus the inclusion of a

linear regression model, as it is easy to apply and in-

terpret, would give a reasonable baseline accuracy if

the more complex Random Forest model did not give

suitable results.

Pickle ﬁles were created for each direction of each

route, containing a dictionary of all stop pairs and

their associated trained model. These were then in-

corporated into the journey planner feature, through

which user inputs could be processed and a prediction

delivered.

https://scikit-learn.org/stable/

GISTAM 2020 - 6th International Conference on Geographical Information Systems Theory, Applications and Management

4 EVALUATION

The focus of this section is on evaluating the perfor-

mance of the machine learning algorithm using the

selected features and segmentation approach. While

the application predicts travel times for over 100 bus

routes, the evaluation focuses on a subset of that num-

ber. Evaluation of the accompanying application’s

functionality and interface is left for future work.

4.1 Predictive Modelling for Accuracy

Multiple strategies were employed to improve the ma-

chine learning model to make it more accurate and

efﬁcient. These included adding more data, remov-

ing outliers and handling missing values, examining

the correlation of features to select the most effec-

tive, tuning the algorithm by optimising the number

of decision trees, and engineering features to examine

their effect on the model. A cross validation approach

was adopted for evaluations. The sample of data we

had available was split into a test set and a training

set. The training set was used to build the classiﬁer

and the test set was used to determine the accuracy

of these classiﬁcation models. The models were pri-

marily evaluated using Mean Average Error (MAE)

and R

because they are robust and a simple metrics

to apply and interpret.

To evaluate the models, three sample route cate-

gories were chosen - ‘urban’ and ‘rural’ routes (based

on travel through a deﬁned area in city centre), and

whether a route was ‘frequent’ (if there is a period of

a headway of ﬁfteen minutes or less on weekdays).

Figure 12 shows the accuracy of the Random Forest

models for six routes. The ‘Whole Journey Predic-

tion’ column shows the metrics of an entire journey,

trained on ‘actualDuration’ from the ‘Trips’ data. The

‘Timetable Prediction’ contains the static estimated

duration of a whole journey on this route, taken from

Dublin Bus’s timetables. The ‘Segmented Journey’

adds the predictions of all stop pair segments’ dura-

tion for the journey.

Figure 12: Accuracy of Six Representative Random Forest

Models.

Figure 12 displays the MAE which shows that on

average, the difference between the predicted value

and the observed value. The ﬁgure also includes

RMSE, which is Root Mean Squared Error. It pro-

vides a similar metric to MAE, except instead of a

linear equation, it uses a quadratic formula to weight

large differences more, making it a particularly useful

metric when large errors are undesirable. The magni-

tude of the difference between the MAE and RMSE

give an indication of the variance in error.

4.1.1 Results and Critique of Predictive Model

Evaluation

The most accurate predictions were the full jour-

ney time predictions of the Random Forest models.

In comparison with the static timetabled results, the

‘whole journey’ predictions perform signiﬁcantly bet-

ter across all routes. These models still perform well

even when there are few rows of data, as in the case

of route 31B. In contrast, the aggregated journey seg-

ments are often less accurate than the timetabled pre-

dictions. Although under-performing when compared

to the ‘whole journey’ predictions, the ‘segmented

journey’ predictions do produce better metrics when

they have access to more data, as in the case of the

routes 46A and the 4. However, it is important to note

that the ‘segmented journey’ predictions divide each

row into stop pairs, so models are based on far less

data than ‘whole journey’ models. It is clear from

the table that accuracy steadily declines as less data is

available, with the exception of route 31B.

Although the aggregated ‘segmented journey’ pre-

dictions were the least accurate, the approach offers

the most ﬂexible planning and is likely to be more

accurate in planning a speciﬁc journey. The Dublin

Bus timetable does offer a breakdown of segments,

but these are not detailed (for example, the 46A route

has 61 stops, but breaks this into only 5 segments for

estimations). On a typical journey, most users will

not travel from the ﬁrst stop of a bus line to its last.

In addition, the general data trend is that segmented

journey accuracy improves with a larger amount of

data.

5 CONCLUSION

While the predictive model that was implemented was

not as accurate as hoped, it provides insight on the

best approaches to a problem like this. With limited

access to hardware, it was not possible to utilize all

the data that was provided, instead having to sam-

ple from the dataset. The results of the valuation

Design and Development of an Application for Predicting Bus Travel Times using a Segmentation Approach

show that providing more data has the potential to

improve the results. Furthermore, while segmenting

the journey into stop-pairs seemed like the smartest

way to approach the problem, especially with the

bus system in Dublin, when data was sparse, it pro-

vided worse accuracy than the static timetable. In

cases where there is more data in the dataset, the seg-

mented journey model does perform better than the

static timetable; however, the whole journey model

always outperforms the static timetable. In cases with

less data, the whole journey model can be about 1000

seconds more accurate. With the right hardware and

the ability to analyse the entire dataset, the segmented

journey model may have returned better results. The

ﬂexibility that the segmented model would provide

would be a good ﬁt for a journey planning applica-

tion, and any opportunity to incorporate that into a

public transportation application would allow users to

receive better results when trying to estimate the jour-

ney time for their bus routes. The next steps of the

work are to increase the amount of data to train the

model for the segmentation approach. User testing of

the application will also be carried out the ascertain

the effectiveness of the interface and associated tools

described in Section 3.

ACKNOWLEDGEMENTS

The team would like to thank the National Transport

Authority in Ireland for providing the data used to

train and test the models in this research.

REFERENCES

Breiman, L. (2001). Random forests. Machine learning,

45(1):5–32.

Brilhante, I. R., Macedo, J. A., Nardini, F. M., Perego, R.,

and Renso, C. (2015). On planning sightseeing tours

with tripbuilder. Information Processing & Manage-

ment, 51(2):1–15.

Celan, M., . L. M. (2017). Bus arrival time prediction based

on network model. Procedia Computer Science, pages

138–145.

Chang, H., Park, D., Lee, S., Lee, H., and Baek, S. (2010).

Dynamic multi-interval bus travel time prediction us-

ing bus transit data. Transportmetrica, 6(1):19–38.

Crist

obal, T., Padr

on, G., Quesada-Arencibia, A., Alay

on,

F., de Blasio, G., and Garc

ıa, C. R. (2019). Bus travel

time prediction model based on proﬁle similarity. Sen-

sors, 19(13):2869.

Gal, A., Mandelbaum, A., Schnitzler, F., Senderovich, A.,

and Weidlich, M. (2017). Traveling time prediction in

scheduled transportation with journey segments. In-

formation Systems, 64:266–280.

Gkiotsalitis, K. and Cats, O. (2017). Exact optimization

of bus frequency settings considering demand and trip

time variations. In 96th Transportation research board

annual meeting.

Huang, R. and Peng, Z.-R. (2002). Schedule-based path-

ﬁnding algorithms for transit trip-planning systems.

Transportation Research Record, 1783(1):142–148.

Jariyasunant, J., Work, D. B., Kerkez, B., Sengupta, R.,

Glaser, S., and Bayen, A. (2011). Mobile transit

trip planning with real-time data. Technical report,

UC Berkeley: University of California Transportation

Center.

Jeong, R. and Rilett, R. (2004). Bus arrival time prediction

using artiﬁcial neural network model. In Proceedings.

The 7th International IEEE Conference on Intelligent

Transportation Systems (IEEE Cat. No.04TH8749),

pages 988–993.

Jiang, F. (2017). Bus transit time prediction using gps data

with artiﬁcial neural networks. https://bit.ly/2wRujmI.

Kroes, E. and Daly, A. (2018). The economic value of

timetable changes. Transportation Research Proce-

dia, 31:3–17.

Lee, W.-C., Si, W., Chen, L.-J., and Chen, M. C. (2012). A

new framework for bus travel time prediction based on

historical trajectories. Proceedings of the 20th Inter-

national Conference on Advances in Geographic In-

formation Systems.

Liebig, T., Piatkowski, N., Bockermann, C., and Morik, K.

(2017). Dynamic route planning with real-time trafﬁc

predictions. Information Systems, 64:258–265.

Lin, H.-E., Zito, R., Taylor, M., et al. (2005). A review

of travel-time prediction in transport and logistics. In

Proceedings of the Eastern Asia Society for trans-

portation studies, volume 5, pages 1433–1448.

Mazloumi, E., Rose, G., Currie, G., and Moridpour, S.

(2011). Prediction intervals to account for uncertain-

ties in neural network predictions: Methodology and

application in bus travel time prediction. Engineering

Applications of Artiﬁcial Intelligence, 24(3):534–542.

Peek, G.-J. and van Hagen, M. (2002). Creating synergy

in and around stations. TRB 81stannualmeeting CD-

ROM. Washington, DC.

Peng, Z., Jiang, Y., Yang, X., Zhao, Z., Zhang, L., and

Wang, Y. (2018). Bus arrival time prediction based on

pca-ga-svm. Neural Network World, 28(1):87–104.

Petersen, N. C., Rodrigues, F., and Pereira, F. C. (2019).

Multi-output bus travel time prediction with convolu-

tional lstm neural network. ArXiv, abs/, 1903:02791.

GISTAM 2020 - 6th International Conference on Geographical Information Systems Theory, Applications and Management