Transport  Modelling  in  the  Age  of  Big   Data  

 
Cuauhtémoc  Anda  
Pieter  Fourie  
Alexander  Erath  
 
 
 
 
 
 
 
 
Work  Report   June   2016  
 
 
 
 

   
 Transport Modelling in the Age of Big Data June 2016
 
 
New Big Data sources such as mobile phone call data records, smart card data and geo-coded
social media records allow to observe and understand mobility behaviour on an unprecedented
level of detail. Despite the availability of such new Big Data sources, transport demand models
used in planning practice still, almost exclusively, are based on conventional data such as travel
diary surveys and population census. This literature review brings together recent advances in
harnessing Big Data sources to understand travel behaviour and informing travel demand
models that allow to compute what-if scenarios. From trip identification to activity inference, we
review and analyse the existing data-mining methods that enable these opportunistically collected
mobility traces inform transport demand models. We identify that future research should tap on
the potential of probabilistic models as commonly used in data science. Those data mining
approaches are designed to handle the uncertainty of sparse and noisy data as it is the case for
mobile phone data derived mobility traces. In addition, data fusion approaches should be applied
to integrate disparate but related datasets to blend Big Data with more granular information
from travel diaries. In any case, we also acknowledge that sophisticated modelling knowledge
has developed in the domain of transport planning and therefore we strongly advise that still
domain expert knowledge should build the fundament when applying data driven approaches
in transport planning. These new challenges call for a multidisciplinary collaboration between
transport modellers and data scientists.

i
 Transport Modelling in the Age of Big Data June 2016

1   Introduction  
 
 
Before the ubiquitous sensing of human mobility flows became possible through mobile phones,
public transport smart card transactions or Global Positioning System (GPS)-enabled devices, it
was difficult and expensive to generate large scale or even population-wide samples that allow
developing travel demand models. The aim of such models is not only to replicate with relevant
accuracy actual mobility flow but also the application of what-if scenarios to evaluate the impact
of different infrastructure development decisions.

Despite the availability of such new Big Data sources, transport demand models used in planning
practice still, almost exclusively, are based on conventional data such as travel diary surveys
and population census. While the applied statistical models have become more sophisticated as
computation power grew exponentially over the last decades, the most important change being
the evolution from trip- to activity-based models, the basic modelling paradigm remained the
same: mobility travel diary survey that only cover a small sample of the actual population is
being used to synthesise transport flows of a representative population.

New Big Data sources such as mobile phone call data records, smart card data and geo-coded
social media records allow us to observe and understand mobility behaviour on an unprecedented
level of details. But simply observing is not particularly helpful for planning purposes. To allow
for prediction in what-if scenarios, we need to understand and contextualise the information
contained in such Big Data sources to inform models of travel behaviour and adapt them to be
useful in travel demand modelling frameworks.

This literature review brings together recent advances1 in the fields of harnessing Big Data
sources to understand travel behaviour and inform travel demand models that allow to compute
what-if scenarios. To this end, we first provide a primer on the latest advances in transport
demand modelling including the latest agent-based approaches.

Our focus is then on the research that makes use of the relevant Big Data sources and directly ties
in the methodological toolkit of travel demand models, hence excluding methods that primarily
derive real-time analytics from such Big Data sources. The aim of the paper is to provide the
reader with an overview how Big Data already improve the understanding of mobility flows and
has been applied for transport demand models from a methodological angle. From this collection,
we identify the advantages and disadvantage of the various methodologies and its applicability
for being used in predictive transport models. Conclusions drawn from this literature review
include the description of new modelling applications that the new data sources allow but also
the specification of research gaps that need to be overcome to realise them.

1
Mainly from 2010 to the first quarter of 2016

2
 Transport Modelling in the Age of Big Data June 2016

2   Travel  demand  models  and  mobility  data  

 
 
2.1   Transport  Demand  Models  
 
 
Travel demand models have been developed to support decision making by forecasting the
impacts of alternative transportation and land use scenarios. (Castiglione et al., 2015). The
following two approaches for transport demand forecasting can be distinguished: first, aggregated
models in which travel demand is specified as aggregated transport flow between zones, and
second, agent-based models that maintain travel demand at the level of individuals throughout
the model.

2.1.1   The  classical  four  step  model  

 
 
’4-step’ demand modelling was introduced in the 1960s (de Dios Ortúzar and Willumsen, 2011).
Originally specified as trip-based models they aim at predicting the number of trips for different
travel modes and routes taken between any two origin and destination zones. The first trip
generation component estimates the number of trips produced by and attracted to each zone.
The second trip distribution step connects where trips are produced and where they are attracted
to. The third mode choice step determines the travel mode, such as automobile or bus, used for
each trip, while the fourth assignment step predicts the routes used for each trip and hence allows
to model congestion induced traffic delays. Since such delays can influence the mode and route
choice but also location choice behaviour, feedback loops including steps 2, 3 and 4 are usually
introduced. Data requirement for four step modelling includes household travel survey
information, census information, and a representation of the transportation network.

2.1.2   Activity-­based   models  

 
 
Since the early 1990s, activity-based models have been promoted as a superior alternative to
four-step models, avoiding some of the inherent limitations of the latter type of models. To
appreciate the significance of activity-based models, Rasouli and Timmermans (2013) emphasise
that four-step models are aggregate in nature – the unit of measurement is not an individual, but
rather the number of trips emanating from any particular zone. In addition, the four-step model
lacks consistency and congruence on how behavioural parameters are used across various
submodels. Furthermore, the assumption of independency between the four modelling steps

3
Transport Modelling in the Age of Big Data June 2016

are regularly cited as key shortcomings when it comes to the the evaluation of travel demand
management policies such as mobility pricing.

The fundamental principle of activity-based models is the understanding that travel is ultimately
derived from the necessity to participate in activities. The aim of activity-based models is to
predict for each individual the number, sequence and type of the activities are conducted over a
certain time-period subject to a set of spatial, temporal and resources constraints. However, while
activity-based models allow to generate spatially and temporally disaggregate description of
travel demand, for route choice and traffic simulation this travel demand is often aggregated again
to so-called origin-destination-matrices that describe how many trips are conducted between
any two OD-pairs. This restriction was originally due to the lack of simulation models that are
suitable to simulate traffic for a relevant spatial extent, i.e. entire cities or regions and across an
entire day, but still apply today due to the computational requirements of agent-based transport
simulation.

In addition to four step model data requirements, activity based models do require one additional
type of input, a ’synthetic population’ at the level of individual households and persons that is
representative for the actual population of the area of interest. This ’synthetic population’
includes a set of socio-demographic attributes which are then used for travel demand modelling
processes. Moreover, for every agent in the synthetic population, a fully descriptive daily activity
plan, including locations of daily activities such as work or education needs to be derived.

Agent-­based  transport  models   Agent-based transport models for strategic transport planning
usually derive travel demand from activity-based modelling approaches but employ microscopic
and completely time-dynamic traffic simulation of each agent’s individual demand based on
system constraints given by the transport network and its attributes. (Balmer, Axhausen and
Nagel, 2006).

While the original development of Transims Smith et al. (1995) as the first large-scale agent-based
transport simulator clearly focused on replacing aggregated transport assignment methods, later
implementations of TRANSIM and more recent developments of agent-based model such as
MATSim (Horni et al., 2016), SimMobility (Adnan et al., 2016), SimAGENT (Goulias et
al., 2012), integrate to different degrees also mode, time, destination and activity scheduling
processes into a single consistent modelling framework. Such an integrated modelling framework
overcomes that travel demand can in disaggregated form throughout the whole modelling process.
Besides the enhanced behavioural consistency, this also allows the modelling and analysis of
modern travel demand management tools such as time- and demand-dependent pricing and new
forms of mobility such as shared and autonomous vehicles.

4
 Transport Modelling in the Age of Big Data June 2016

Multi agent-based modelling is built upon a large scale of autonomous agents which perform
their own decisions, interact with one another and with the environment. For each agent, an
initial daily activity plan is assigned as a precise description of the activities’ location, its
durations, start and end time, and the trips connecting two activities, including mode and route.

Among the several agent-based transport models that are currently under continued development,
MATSim takes a special role and can be considered the the currently most widely applied model.
It can integrate a wide range of decision dimensions in a co-evolutionary learning loop, but due to
its modular framework, it can also be used for traffic simulation only and integrated with other
activity-based travel demand models.

In MATSim, a day is simulated multiple times and after each iteration a fraction of the agents is
allowed to modify their plans (i.e. mutation/crossover phase). For instance, they can change
their departure time, the travel mode of a sub-tour, location of a given type of activity, among
others. At the end of each simulated day, the utility of the day is measured for each agent using
a scoring function that rewards agents for performing activities, while penalising them for
travelling, transferring between transport modes, waiting at transit stops and arriving late for
activities, etc. In such way, agents seek to improve their utility over iterations until the system
reaches an equilibrium where the generalised utility can not be longer improved (i.e. a steady-
state is reached) (Balmer et al., 2009).

2.2   Big  Data  describing  mobility  

 
 
Precise geo-referenced location data represents a large and growing subset of Big Data as
mobile devices and location-sensing technologies become ubiquitous. For the purpose of
transportation planning, and accordingly to disaggregate activity-based approaches, the aim of
this literature review paper will be limited only to mobility data sources emerging from
individuals. We are interested in the digital trace left by individuals because it can provide us
with more accurate and interesting insights on mobility patterns. Hence, out of the scope will
remain mobility information generated by infrastructure sensors that document traffic flows at
certain cross-sections (e.g. loop detectors, video vehicle detection systems, ERP systems).

Due to its wide-coverage in urban settings, the main focus of the survey review will be firstly on
data generated through Smart Card Automatic Fare Collection (SC-AFC) systems, and mobile
phone networks. Both of them can be classified as large-scale opportunistic human mobility
sensors, which are able to provide insights on urban dynamics and human activities at an
unprecedented scale and level of detail. Plus, the advantage that no additional infrastructure is
needed to extract mobility information, since they were designed to collect for public transport

5
 Transport Modelling in the Age of Big Data June 2016

fare and allow mobile communication network usage.

In addition, we will also cover research working with GPS data. Its high-resolution and accuracy
has allowed the creation of preprocessing and data mining techniques as well as inference models
that go beyond SC-AFC and mobile phone data. Plus, the recent drop in cost that allowed the
widespread use of GPS sensors in mobile phones and other devices, outlooks an obvious potential
for wide-spread application as urban mobility sensors in the future. Other supplementary datasets
being reviewed are Points of Interest (POI) information, Census and Surveys, and Land use
information. As commented on Calabrese et al. (2015), the purposes of the supplementary
datasets are three-fold. 1) to validate findings extracted from the analysis of large-scale human
mobility sensors. 2) to define scaling factors to extend results to the overall population. 3) to
augment information about urban space, to be able to extract higher level patterns.

2.2.1   Mobile  phone   data  

 
 
Purpose: Large-Scale Human Mobility Sensor

From the ubiquitous computing devices, mobile phones have the highest levels of penetration
rate. While conventional mobile phones usually only sporadically exchange information with
cell-towers, the widespread use of smart phones - the level of penetration rate for smart phones
in Singapore reaches 88% of the population2 - provide appealing new opportunities to inform
travel demand model. Smart phones not only exchange data much more frequently with the
mobile network provider allowing more continuous tracking, they also carry a series of additional
sensors provide such as GPS that can be used to better understand mobility patterns. For instance,
Airsage3 is a company that process mobile phone data for transport planning applications.

Mobile  phone  networks   Whether it is GSM, CDMA or LTE, mobile phone networks require
regular and frequent handshakes (i.e. pings) between mobile phone devices and cellular
communication antennas. In order to provide service to the users, mobile phone networks are
constantly and frequently determining the location of the mobile phone devices even if it they
are simply on standby. The user’s location is calculated by determining the location of the cell
antenna closest to the handset. This results in a precision equal to the size of the cell antenna
coverage, which can range among few hundred metres in urban areas.

In order to understand the events that generate the user location updates we need first to know
how the mobile phone network is constituted. The service coverage area of a given mobile phone

2
The Connected Consumer Survey 2014/2015. https://www.consumerbarometer.com/
3
www.airsage.com

6
Transport Modelling in the Age of Big Data June 2016

network is divided into smaller areas of hexagonal shape, referred to as cells. Each cell is the
area in which one can communicate with a certain base station antenna (also referred as cell
tower). A Location Area (LA) is a geographic area covered by base stations antennas belonging
to the same group. For users to be reached wherever they are in the network coverage area, we
can divide the update location procedures in network-triggered and event-triggered.

Network-triggered location updates occur when a mobile phone is,

1.   being switched on and connects to the cellular network.

2.   involved in a call and moves between two different cell areas (i.e. handover)
3.   on standby and moves into a cell which belongs to a new LA.
4.   polled by the network as its associated timer has come to an end (i.e. periodic location
update, usually every 2 h)

Event-triggered updates happen in the following situations.

1.   When a call is placed or received

2.   When Short Message Service is used (sending and receiving).
3.   When the user connects to the Internet (e.g. to browse the web, or through email
periodically server check)

Most studies found in literature that use mobile network data analysing individual mobility
patterns from mobile phone data make use of Call Detail Records, which is a subset from the
mobile network data used for billing purposes, but also include event-triggered updates.
However, there are also a series of studies that make use of the full spectrum of mobile network
data, including both the event-triggered and network-triggered updates.

Moreover, precision of cell antenna location data can be improved between a dozen to hundred
meters when information is triangulated with signals from other cell antennas. (International
Transport Forum, 2015). By using the Timing Advance, which is a value that corresponds to the
length of time a signal takes to reach the cell tower from a mobile phone, information can be
triangulated from different cell antennas to have a more accurate estimation on the user’s
location. Other techniques used are based on the signal strength received by the mobile phone
using known irradiation diagrams and propagation models for localisation.

7
 Transport Modelling in the Age of Big Data June 2016

2.2.2   Smart  card   data  

 
 
Purpose: Large-Scale Human Mobility Sensor

Smart Card Automatic Fare Collection (SC-AFC) systems are built on radio-frequency
identification (RFID) technology. The objective of RFID is to use radio waves to exchange data
between a reader and an electronic tag for the purpose of identification and transaction operations.
Specifically, SCAFC systems use NFC (Near Field Communication) technology which ensures
secure short range RFID transactions. Since 1990, the use of smart card has become significant
in many sectors since it is perceived as a secure method of user validation and fare payment.
(Trépanier et al., 2007).

For the case of public transportation, commonly, users have to tap in their smart card onto the
reader device at the entrance of buses or metro stations and tap out at the alighting bus stop or
metro station. Besides revenue collection, large quantities of individual detailed information are
collected such as boarding times, boarding stations, alighting times, alighting stations, vehicle
identification. This represents a huge potential in better understanding travel behaviour and
improving current transport systems.

2.2.3   GPS  
 
 
Purpose: Model Transferability

GPS information allows the collection of a detailed spatio-temporal trace describing the mobility
of an individual. All smartphones are usually equipped with GPS sensors and also some
conventional mobile phones nowadays include it. In order to calculate position, GPS uses
information of at least four satellites in a method called trilateration. In open areas accuracy can
be achieved up to 5 meters but it degrades, however, in areas where GPS signals are impaired by
tall buildings or trees and inside of buildings.

Assisted-GPS (A-GPS) increases location accuracy in urban areas by combining GPS location
signals with cellular location data providing an under 10 meters precision. Similarly, other
forms of hybridised GPS location systems can include the use of Wi-Fi network signals through
the tracking of media access control addresses (MAC addresses) within a network of Wi-Fi
routers.

8
 Transport Modelling in the Age of Big Data June 2016

2.2.4   POIs  
 
 
Purpose: Urban Space Augmentation

Points of Interests (POIs) are a list of business and important places to visit in a city, including their
name, classification and location. There are many possible different sources: Yellow Pages, Google
Places, Yahoo PlaceFinder, which might provide different information. For instance, in Google
Places we can find in addition opening times, reviews and hourly-estimates on the crowdedness of
a place. Furthermore, opportunistic POI datasets can be derived from crowdsourcing platforms and
social networks (e.g. Foursquare, Flickr, Twitter, etc.). The importance of these datasets is the
potential to serve as complementary datasets to augment urban space information and thus, improve
activity and places inference estimations.

2.2.5   Census  and   Surveys  

 
 
Purpose: Validation and Scaling Factors

Census and surveys provide datasets related to very different areas: demography, health, education
government and security, communication and transport, etc. Such datasets can be used to: 1)
validate home and working areas, 2) validate city patterns such as hotspots, commuting, traffic
flows, 3) validate land use. The main advantage is the very refined spatial resolution which is
often the census block. The main disadvantages are that they are updated usually only every 5 to
10 years.

Other use of Census information is that it provides the means to perform scaling expansion from
information derived by large-scale human mobility sensors.

While the use of census and other conventional survey data has a long tradition in travel demand
modelling and is well documented, the review of the related work does not fit the scope of this
review. However, since census and survey data is still relevant to enrich and mobile phone data
through data fusion, this review brings together pioneering work in those areas.

9
 Transport Modelling in the Age of Big Data June 2016

2.2.6   Land   Use  

 
 
Purpose: Urban Space Augmentation

Land use datasets offer access to various information that allows to characterise an area based
on its planned and effective land use. Such land use data usually specifies for each plot the
designated usage purpose in case of built up environment also the usage intensity. However, as
different authorities use different land use classifications, the developed models are usually
customised to local conditions which restricts the direct transferability between regions. Within
the scope of this review, land use data is of interest due to its potential to impute the purposes of
activities as identified from SC-AFC and mobile phone data.

10
 Transport Modelling in the Age of Big Data June 2016

3   Smart  Card  Data  

 
Smart Card Automated Fare Collection (SC-AFC) systems are used in many public transportation
systems around the world and continue to be adopted by public transport operators. The main
objective of introducing smart cards for public transport is its ability for flexible and secure fare
collection. However, the information generated with any transaction (time and location) has
soon become recognised as a rich data source for transport and urban planning and science. From
public transport ridership analysis to the creation of OD matrices, smart card data offers insights
on urban dynamics and mobility patterns of a city’s public transport. The following chapter
presents the literature on ways to exploit smart card data for transport planning – from individual
trip reconstruction to the estimation of OD matrices and its inclusion to agent-based simulations.

3.1   Individual  Trip  Reconstruction  

 
Implementation of SC-AFC systems varies depending the city and its fare policies. Some cities
like Amsterdam, Sydney, and Singapore charge public transport fares based on the total distance
travelled, regardless if it is on a bus or train. This requires commuters to tap their smart card
when they board and tap it again in the alighting stop or station. Nonetheless, other cities like
London have non-ladder fares in which the fare is identical for the whole line regardless of where
you get on or get off, thus commuters are only required to tap the smart card once. In any of the
cases, to further analyse human mobility the main challenge of mining smart card data is to
reconstruct the individual trips.

3.1.1   Estimating  Alighting  Stops  

 
 
For SC-AFC systems in which it is only required to validate the boarding location, the first step is
to estimate the alighting stop or station. Generally, to infer the alighting stop the Trip-Chaining
algorithm is used based on two explicit assumptions by Barry, Newhouser, Rahbee and Sayeda
(2002). The first one states that after a trip, users will return to the destination of the previous
trip station; the second one, that at the end of a day, users will return to the station where they
boarded for the first trip of that same day.

Several efforts have been done to improve the original idea by Barry et al. (2002). Zhao et al.
(2007) expanded the idea to rail-to-bus sequences. Trépanier, Tranchant and Chapleau (2007)
incorporated the possibility of looking at the next day, even observing weekly travel patterns to

11
 Transport Modelling in the Age of Big Data June 2016

complete missing information for the bus system of Gatineau in Quebec. Munizaga and Palma
(2012) proposed a multimodal public transport methodology using time constraints instead of
distance constraint. For these studies, the reported success varies from 66% to a 80% of the
individual trips reconstructed.

Furthermore, a different approach to reconstruct individual trips from smart card data which is
based on a semi-supervised learning algorithm has been presented by Yuan, Wang, Zhang, Xie
and Sun (2013). They proposed an integrated learning method in which they align the monetary,
geospatial and temporal spaces to extrapolate a series of critical domain specific constraints.
They incorporate those constraints in a semi-supervised conditional random field algorithm to
infer the exact boarding and alighting stop even if there exist records on trips with unknown
boarding and alighting information. Given only 10% trips with known alighting/boarding stops,
they inferred more than 78% alighting and boarding stops from trips with missing information.
The relevance of the work is not only the reconstruction of origin-only labelled trips, but a
systematic way to recover individual mobility history from urban scale smart card transactions.
This can be helpful as a pre-processing stage for later analysis or incorporation to transport
demand models.

3.1.2   Stages,  Trips  and  ODs  

 
 
After the alighting location is known, the second step in the individual trip reconstruction is to
infer whether the alighting location is the final destination (thus the trip is complete) or if it is
only a stage of a multi-stage trip (i.e. a transfer). The common approach to identify stages is with
a time-based rule. For instance, Munizaga and Palma (2012) used a 30 minute rule. If a person
stays longer than 30 minutes in a particular point, then it is said to be the destination. For the
case of London, Seaborn, Attanucci and Wilson (2009) recommended elapsed time thresholds
depending on the transfer type — 20 min for underground-bus, 35 min for bus-to-underground,
and 45 min for bus-to-bus.

Having information only on smart card data draws a limitation in identifying the spatio-temporal
dimension of an individual performing an activity, since public transport is not exclusively used
for all trips throughout a day. Chakirov and Erath (2012) describes the limitation with the
concept of public transport trip consistency, in which consistency means that a person who
arrived to the activity location by public transport, has to leave it after ending the activity also by
public transport. Although smart card data do not record any other means of transport except
from public transport, the most obvious cases of inconsistency can be identified by analysing
the distances between the alighting location of the last journey and the boarding location of the
following journey. This allows to identify if other means of transport such as taxi, car or walking

12
 Transport Modelling in the Age of Big Data June 2016

must have been used in between.

For instance, on a typical workday in Singapore, Chakirov and Erath (2012) found that from
persons with more than one journey recorded in the smart card data, 90% of the journeys, start
less than 1 km away from the previous alighting location. This indicates firstly, that the majority
of public transport users don’t switch to other modes of transport between public transport
journeys and therefore have consistent journey chains. And secondly, that with some degree of
uncertainty an area can be limited for possible activity locations.

Once individual trips have been reconstructed to the point of knowing boarding locations and
final destinations, a possible application is the calculation of a public transport OD matrix. It is
important to take into account trips which could not be able to reconstruct. For this situation a
typical solution is to build expansion factors. The work by Munizaga and Palma (2012) shows
how to build expansion factors for smart card data trips associated with an origin but not with a
destination, and for trips associated with no origin or transaction. For the former ones, it is
assumed that the distribution of trips is the same as that of other trips with the same origin; whilst,
for the latter case, the distribution of trips is associated only by their time disaggregation.

3.1.3   Activity   Identification  

 
 
Public transport consistent trips can be further studied to introduce semantic meaning to the
inferred locations. Devillaine, Munizaga and Trépanier (2012) present a direct rule-based
classification including information on the card type and the temporal attributes of the trips.
Work is assigned for adult cards, for which the activity is longer than 2 hours and the trip before
the activity is not the last of the day. Similarly, study is assigned for student or underage cards,
with activities longer than 5 hours and the activity not being the last one of the day. Finally,
home is assigned if the trip after the activity was the last of the day, and others was assigned to
the rest of activities. The criteria used represents a generalisation on the working/studying and
staying at home behaviours, which might not be accurate for an important percentage of the
population. Shortcomings of such a rigid classification can be improved by the introduction of a
probabilistic choice models, which can take into consideration land use information as well.

Such is the case of the work by Chakirov and Erath (2012). They proposed a multinomial logit
model with activity duration, activity start time and land use as the the utility variables to match
a discrete-choice space consisting of work activity, home activity and other activity as the target
labels. Utilities for the model were constructed mainly using piecewise linear functions. For the
case of activity duration and start time, the utility function was calibrated using information

13
 Transport Modelling in the Age of Big Data June 2016

from the local travel diary survey, whilst for the land use information stemmed from the urban
planning authority’s Master Plan.

Another probabilistic approach was proposed by Isaacman, Becker, Cáceres, Kobourov,

Martonosi, Rowland and Varshavsky (2011) who built a continuous-space model to detect home
and work locations. They introduced a score function obtained from a logistic regression and
calibrated from a set of trained users. Similarly to Chakirov and Erath (2012) the home and work
location labels were identified mainly by the time factors associated to the events. However, the
main differences between the two probabilistic approaches relies not in whether they have
chosen discrete or continuous space for their methods, but that the calibration process diverges
in using information from other data sources (i.e. household travel survey and land-use) in a
transfer-learning scheme (Chakirov and Erath, 2012) against the traditional learning scheme, in
which calibration comes from a labelled training subset (Isaacman et al., 2011).

After having obtained the home and work locations, a validation process is required to determine
the model’s accuracy. One possible validation process is the one used by Yuan et al. (2013). A
2D Kernel Density Estimation (KDE) was applied to identify hot spots that enabled a comparison
against data derived from travel diary surveys.

3.2   Agent  Based  Transport  Models  and  Simulation  

 
 
State-of-the-art methods for travelling demand models are currently led by multi-agent simulation
frameworks. This new paradigm was put forward to overcome some of the flaws of the classic
four step model, namely to ensure behavioural consistency, include temporal travel dynamics in
a continuous manner and maintain the disaggregate nature of transport demand through out the
modelling process. The disaggregate nature of smart card data represents an appropriate input to
such models. Moreover, by assuming that each unique card ID represents an agent, demand for a
multi-agent based transport scenario can be directly derived from the smart card data.

The work of Bouman (2012) represents a first attempt to implement an agent-based micro-
simulation of public transport for the cities of Amsterdam and Rotterdam. Based only on smart
card data, the main challenge of the work is the generation of the agents’ activity plans. For such
task, they focus on the extraction of commuters’ home-work-home pattern looking at several days
of the same user. Work and home stations are identified as the two most visited stations during
the weekdays, from which home is identified as the most visited station during the weekend.
For smart cards ids whose data was not suitable to be fitted with such a pattern, but at least an
activity chain could be reconstructed for one particular day, travel demand was specified by
introducing dummy activities at the intermediate stations of the tour. Finally, for highly irregular
travel patterns, a new agent for each of the remaining trips was generated.

14
Transport Modelling in the Age of Big Data June 2016

The synthetic-population generation process suffers from different limitations that mainly stem
from the various assumptions imposed in the modelling process. Opportunities for future research
can be identified in a more accurate and validated representation of the actual travel demand
including the inference of trip purposes and socio-demographic characteristics of the traveller.
To this end, we consider the long observation periods of smart card data as an opportunity to
apply modern data mining techniques to infer additional information. Along this line, (Bouman
et al., 2013) for example explored how the concept of eigenbehaviours (Eagle and Pentland,
2009) can be applied to derive spatio-temporal patterns.

Another challenge when using smart card data for simulation is to model potential interactions of
public transport vehicles with other transport modes (i.e. cars). A recent work by Fourie, Erath,
Ordóñez Medina, Chakirov and Axhausen (2016) develops a simplified agent-based transport
simulation for Singapore’s public transport. In contrast to Bouman (2012), the interaction with
private vehicles was accounted by introducing a stochastic model of the speed of buses between
public transport stops and bus dwell time behaviour at stops. This allows not only to improve the
simulation time substantially, but also to predict the operational stability of alternative public
transport schedules, making simulations of system-wide network redesigns possible. For this
purpose, they adapted three inputs to the MATSim environment.

Firstly, a reconstruction of bus trajectories from smart card data was developed. Given all boarding
and alighting transactions of bus users, the position in space and time of the corresponding buses
was estimated. They imputed the time it takes for a bus to travel between bus stops locations by
grouping its transactions at each stop into sets that represent bus dwell operations. Then, from
the reconstructed bus trajectories they determined the number of services and the time when the
services start for every bus line in Singapore. For the particular case of train services, they took
the start times from the Google Transit Feed Specification (GTFS) since especially during peak
hours train trajectories cannot reliably inferred from public transport smart card data Sun et al.
(2012).

Secondly, they needed to generate activity plans for each agent in the simulation. To that end,
they established a 25 min threshold to identify the final alighting location of each multistage
trip to not split journeys at transfer points. Since smart card records only document boarding
times but not when a person actually arrived at the bus stop: given that average headway for
most bus services is 10 minutes or shorter, they assumed a uniform arrival time distributions
and randomly drew the actual arriving time from the bus stop with the only parameter being the
corresponding headway between consecutive services of the specified line.

15
Transport Modelling in the Age of Big Data June 2016

Lastly, as only public transport vehicles were simulated, a simplified network offered the
opportunity to lower the computational demands of the simulation substantially. Instead of the
MATSim queue model, a stochastic travel time model was introduced to model travel times
between two subsequent stops. The model was fitted based on a multinomial regression model
assuming that stop-to-stop follow a normal distribution Fourie (2014). As shown by Sarlas and
Axhausen (2015) the parameters that determined the speed of vehicles in a network link were
related not only to the level of demand on the link (taken from smart card data), but also to the
topographical information contained in the network description. To account for dwell time
variability in the simulation framework they included the model presented by Sun, Tirachini,
Axhausen, Erath and Lee (2013).

As a case study to showcase the abilities of the model, they simulated the impact when splitting
one of the longest bus lines in Singapore. The results suggested that incidences of bus bunching
can be significantly reduced during the morning peak hour, and that headway reliability is also
improved considerably.

The work of Fourie et al. (2016) shows one of the possible integrations of big data algorithms
within an agent-based transport modelling framework. They enhanced the simulation results of
a complex model by introducing what is called in the computer science literature as a machine
learning surrogate. They substitute a compartment of the whole model with the behaviour
obtained from the statistics of the smart card data. The results not only represent a more accurate
representation of the real world, but also an improvement of the overall computation time.

However, still several limitations have to be addressed such as implementing the passing behaviour
in the queue simulation, the reconstruction of train trajectories, a better representation of walking,
waiting and transfer activities to better represent route and mode choice preferences. In addition,
public transport smart card data obviously do not contain any information of motorised travel
demand and active mobility. To tackle the aforementioned problems, additional datasets can be
included.

16
 Transport Modelling in the Age of Big Data June 2016

4   Mobile  Phone  Data  

 
Data derived from mobile phones network transactions constitute a potential source of information
for models of daily activities and transportation. In contrast with household survey interviews,
mobile phone data offers large sample sizes and long observation periods at negligible costs.
However, one has to overcome the challenges of processing mobile phone traces for trip
reconstruction — the information contain is such data streams is low in spatial resolution and
sparse in time. Concretely, the precision on the location estimates depends on the cell tower
distribution in a given area, whereas the frequency on the location updates is characterised by
each user’s usage.

Moreover, for the different studies found, the quality of the data also depends on the type of
dataset provided by the mobile network operator. While some studies work with the full
spectrum of mobile phone management signals (e.g. location area (LA) handles, device updates),
others work with the Call Detail Records (CDR) subset, or with already preprocessed data (e.g.
triangulated location estimates). Thus, the general challenge in using mobile phone data is how
to robustly extract people’s trip sequences from sparse and noisy measurements and enrich the
extracted trips with semantic meaning (i.e. trip purpose) (Widhalm et al., 2015).

In the following, we present a series of studies that aimed to reconstruct individual trips from
mobile phone data in order to extract mobility patterns. Thereby, we focus on approaches that
allow the generation of OD-matrices and applications with potential to be adopted for agent-
based simulations. For this purpose, we have divided the relevant literature based on their
methodological approach and their scope. The first group of studies introduced resembles the
traditional trip-based approach, a second group focuses on extracting stay locations from noisy
mobile phone traces, and the third group attempts to infer activities performed at the extracted
locations. At the end of the chapter, we finalise presenting the studies that have been made using
mobile phone data specifically for the generation of agent-based simulations.

4.1   Trip-­based  OD  reconstruction  

 
 
The first approaches found in literature that aim at understanding mobility flows based on
mobile phone refer to attempts to generate OD matrices. Caceres, Wideberg and Benitez (2007)
performed a pilot study using CDR simulated data to match origin and destination areas with the
LAs used for the management of the mobile phone network. The idea behind the study was to
demonstrate the correlation between the movements of anonymous mobile phones and vehicle
displacements.

17
 Transport Modelling in the Age of Big Data June 2016

This idea was then formalised by Wang, Hunter, Bayen, Schechtner and González (2012) with
the definition of the transient OD matrix. The concept recognises that even if segments of the
trip are unobserved in the CDR (e.g. real origin and destination locations) due to mobile phone
inactivity, still a large portion of the actual ODs are retain and can be use to analyse road usage
patterns. The transient OD matrices are constructed by simply counting trips for each pair of
consecutive calls made within the same hour from two different towers, and then the OD trips
are assigned to the road network by a shortest path algorithm.

Similarly, Iqbal, Choudhury, Wang and González (2014) demonstrate the development of OD
matrices using CDRs from Dhaka, Bangladesh, and traffic counts from a video vehicle detection
system. Firstly, tower-to-tower transient OD matrices are generated and then associated with
corresponding nodes of the traffic network converting them to node-to-node transient OD
matrices. Then, the transient OD matrices are scaled up to match the traffic counts. To determine
the scaling factors, an optimisation-based approach is used which minimises the differences
between observed and simulated traffic counts at the points where the traffic counts are available.
Lastly, for the estimation of the final OD matrix they introduce correction factors to account for
the mobile phone market penetration rates and mobile phone usage.

As discussed by Jiang et al. (2015) one of the problems of the trip-based approach is that it can
introduce biases when CDR data are low in spatial resolution. In addition, the former methods
are not able to handle noisy measurements from raw mobile phone traces. In order to avoid
these issues, the approach of the next group of studies presented is based on parsing the
trajectories observed into stay-locations.

4.2   Stay  location-­based  OD  reconstruction  

 
 
The fundamental premise of activity-based travel models is that travel demand derives from
people’s needs and desires to participate in activities. In the following, we present a set of
methodologies that have been proposed in literature to identify activity locations from mobile
phone data and therefore are potentially relevant also for building activity-based transport demand
models. Activity locations are identified by filtering out passing by points, and estimating arrival
and departure times from raw mobile phone data. In contrast with trip-based models, individual
trips are obtained from the flows between the identified stay-locations. However, in order to
obtain such information, sophisticated mobile phone data processing algorithms are required.

Additional to a lower spatial resolution in comparison with GPS traces, mobile phone data
suffers from a phenomenon called supersonic jumps or signal jumps (i.e. outliers). These are
events that suddenly occur kilometres away within a short period of time. Although such jumps

18
Transport Modelling in the Age of Big Data June 2016

usually are system inherent data noise, some jumps might be triggered by external mechanisms
in aims to protecting the privacy of the users (Horn et al., 2014). To use mobile phone data for
accurate traffic modelling, these shortcomings must be considered in order to derive realistic
trajectories.

Temporal-­based   clustering   Schlaich, Otterstätter and Friedrich (2010) work with Location
Area (LA) updates from a region of southwest Germany. The algorithm proposed is built on
the principle that if a user remains a considerably longer time in a location area than the time
required for directly traversing the area, the user potentially starts or ends a trip in the respective
location area. For this purpose they suggested a 60 min rule in which if the time period between
the first login and the last logout of a multiple visited LA is 60 min or more, then it is considered
to be a stay location. In addition, as a strategy to deal with signal jumps, they calculated a
jumpiness factor and deleted user entries that exceeded a given threshold.

Certainly, the approach suffers from several limitations due to the resolution of the trips extracted
being at the broad LA level and not the cell-tower area level. For instance, in their preprocessing
step, the decision to delete consecutive data points from users with the same location areas,
disables them to estimate arrival times and activity durations. Also, since their method requires
a minimum of three LAs, they deleted users that show records of less than three different LAs.
Hence, a more robust methodology would be desirable that can handle noisy signals and outlier
points without the need to delete any entries, and that can be able to estimate trips at the cell-tower
resolution.

Distance-­based  clustering      Calabrese, Lorenzo, Liu and Ratti (2011) proposed a method to
identify trips at the cell-tower level based on Call Detail Records (CDRs) generated from phone
calls, messages and internet usage. They also included a strategy to handle noisy traces
employing different clustering techniques. For the preprocessing step they first characterised
the individual calling activity and verified that it was frequent enough to allow monitoring the
user’s movement over time with a fine enough resolution. Then, they applied a low-pass filter
with a 10-minute resampling rate and a clustering technique to identify minor oscillations
around a common location. As for the extraction of stay-points, they performed a distance-based
clustering to fuse points within a 1 km area. The centroid of the cluster was defined to be a
virtual location and in a final step, individual trips were reconstructed by connecting the paths
from the identified virtual locations. However, the methodology lacks to robustly filter out
passing by events.

19
Transport Modelling in the Age of Big Data June 2016

Frequency-­based   clustering   An alternative to identify important places (i.e. stay-locations)

from temporally sparse and spatially coarse CDRs is presented in Isaacman, Becker, Cáceres,
Kobourov, Martonosi, Rowland and Varshavsky (2011). They assume that the most visited cell-
towers are connected to important places in a person’s life. Instead of using temporal or spatial
clustering algorithms to obtain those locations, they use the frequency of cell-tower visits. The
method consists in sorting cell towers based on the total number of days they were contacted and
applied a cluster leader algorithm. After obtaining the cell tower clusters, they used a logistic
regression model trained on the activity behaviour of 18 volunteers. The explanatory variables
included on the model were the number of days during which any cell tower in the cluster was
contacted, the span time between the first and last contact with any cell tower in the cluster, the
working hour events (between 1pm and 5pm) and the home hour events (between 7pm and 7am).
While the proposal performs well with the identification of primary activity locations (e.g. home,
work), it is not designed to extract secondary activity places for which they can be confused with
en-route events.

At this point, we can recapitulate and reformulate the main ideas towards extracting places from
raw mobile phone data according to the data mining pipeline suggested by Jiang et al. (2015).
First, the need to eliminate outlier noise and signal jumps between towers. Secondly, the need to
cluster points that are spatially close and temporally adjacent into a single location. And thirdly,
the need to agglomerate points that are spatially close but not necessarily adjacent in temporal
consecutive sequence, since we are interested in the unique stay locations that a user frequents.
In addition, an estimation on activity start times and durations is also needed. The following
studies represent the latest endeavour (from 2013 to 2015) to mine location points from mobile
phone data.

Preprocessing  techniques  For the first goal, Horn, Klampfl, Cik and Reiter (2014) perform an
evaluation on three different types of filters to detect outliers on mobile phone traces: a
Recursive Naive Filter, a Recursive Look-Ahead Filter, and a Kalman Filter. On the one hand,
the first two basically act as low-pass filters (Calabrese et al., 2011). They smooth out large
positioning errors by introducing an upper bound constraint on the travel speed. Hence, the
speed is calculated for each each consecutive pair of points (Recursive Naive Filter), or each
triad of points (Recursive Look-Ahead Filter) and compared to a certain threshold. On the other
hand, the Kalman Filter is a probabilistic approach that reconstructs the trajectory. The results
demonstrated that the Recursive Look-Ahead Filter performed better as it eliminated the outlier
points, and in addition maintained the accuracy of the trajectories. Although the Kalman Filter
also eliminated the outlier points, the trajectories lose accuracy. However, the satisfactory results
of Ficek and Kencl (2012) to extend the spatial resolution of the Reality Mining Dataset (Eagle
and , Sandy) using a Gaussian Mixture Model, suggest that given the low-resolution of CDRs,
more complex probabilistic filters are needed in order to outperform the naïve approaches.

20
Transport Modelling in the Age of Big Data June 2016

Time-­distance  clustering   For the second and third goals (i.e. location extraction), Jiang et al.
(2013, 2015), Toole et al. (2015), and Alexander et al. (2015) used time and distance clustering
techniques to filter out passing by points. Firstly, they grouped points that are spatially close by
measuring the distance between two consecutive points and comparing them to a distance
threshold (e.g. roaming distance of 300 m). Then, the clusters obtained are considered to be
potential stays if the time between the first and the last observation in the cluster are separated by
a time greater than a time threshold (e.g. 10 min). Then, the geographic location of the potential
stay is set to be at the centroid of all points within the cluster. Due to noise in locations, multiple
potential stays that are actually the same place may be estimated at a slightly different geographic
coordinate on different observation days. To account for this, a final agglomerative clustering
algorithm is used to consolidate candidate stays to a single semantic location regardless of the
temporal sequence of the records.

Trip  Validation  It is important to verify for the algorithms proposed, that users with more phone
activity do not have systematic differences in travel behaviour. For instance, that there does not
exist a correlation between the number of places detected and the mobile phone usage. Jiang et
al. (2015) segmented users into 5 groups according to the frequency of total number of phone
usage observations. Then they examined for each group the daily travel patterns, including daily
number of trips, and daily number of unique destinations. Finally, they compared the frequency
distributions of both the number or trips and the daily number of unique locations and conclude
that they follow similar patterns.

Activity  start  times  and  durations   After having identified the stay locations, Widhalm et al.
(2015) continue the study by estimating the arrival time as the average between the earliest record
in the arrival activity (i.e. the upper bound of the arrival time), and a lower bound estimate,
calculated as the sum of the latest record at the previous location plus the travel time between
the previous and present location. The travel time was determined as the distance between the
consecutive centroid of clusters divided by an assumed travel speed. The same process was
performed for the expected time of departure, and the activity duration calculated by subtracting
both estimates.

Another alternative to infer arrival/departure times of activities is the one of Alexander et al.
(2015). They proposed to use probability density functions of activity durations derived from
the National Household Travel Survey. They constructed six hourly distributions for weekdays
and weekends and the following trip purposes: home-based work (HBW), home-based other
(HBO), and non-home based (NHB). Then, they randomly generated the departing time within
the time window of observations, using the distribution that corresponded to the day (weekday,
weekend) and the trip purpose (HBW, HBO, NHB).
21
 Transport Modelling in the Age of Big Data June 2016

4.2.1   Expansion   Factors  

 
 
After locations have been extracted from mobile phone data, trips can be generated by connecting
consecutive locations and OD matrices can be potentially determined. However, in order to
project the mobile phone sample to the entire population, expansion factors need to be carefully
calculated, avoiding any socio-economical biases, and correctly transforming the individual trips
recorded from mobile phone data into vehicle flows while taking into account shared rides (and
also trips made by public transport).

Several authors (Jiang et al., 2015; Alexander et al., 2015; Toole et al., 2015) have proposed the
expansion factor calculation by identifying the tower a user is connected to while being at home.
Since the ratio of mobile phone users to the population is not uniform within the research region,
each user is assigned a home census area, and expansion factors are computed for each area by
measuring the ratio of the total population living in the area, and the sum of users whose home-
tower was identified inside the area. General OD matrices are built, and if a certain mode of
transport wants to be considered, the vehicle OD matrix is approximated by weighting the total
number of user trips by the vehicle usage rate in the home census zones. However, the limitation
of the methodology is the generalisation of travel patterns by census areas which might reflect
biases in the vehicle trips ODs.

Alternatively, Zhang, Qin, Dong and Ran (2010) build an expansion factor based on a probabilistic
approach. They firstly looked at the problem on how to avoid multiplication of trips when there
exist more than one mobile phone probe in a vehicle. They calculate a conversion factor from
mobile phone probe flows into equivalent vehicle flows, using the following assumptions: —
1) Mobile phones in close proximity (i.e. the same car) generate signal transition events at
exactly the same time. 2) There is a very small probability that some parallel travelling cars are
crossing at least two LA boundaries at two same timestamps. 3) Within the saturation headway
(typically 2 seconds) there is only one vehicle crossing an LA boundary in each lane. Secondly,
they computed a conditional probability for mobile phone ownership including mobile phone
market penetration data, market shares for a given mobile phone carrier, and age and income
distributions from census data. Finally, for the purpose of projecting to the population vehicle
OD matrix, they calculated the scaling factor through a Horvitz-Thompson estimator, in which
the conversion factor from mobile phone to vehicle flows was included, as well as the posterior
probability of mobile phone ownership.

In contrast with the home-based approach, the model takes into account the conversion from
individual mobile phone flows to vehicle flows (i.e. shared rides) not as an a posteriori step, but
integrated in their model. Plus, the fact of characterising the population mobile phone ownership
by age and income, instead of zones, avoids socio-economical biases more accurately in the
projection process.
22
 Transport Modelling in the Age of Big Data June 2016

4.3   Activity-­based  ABM  

 
So far we have obtained locations, start times and activity durations from raw mobile phone traces.
However, to derive an activity-based travel demand description to be used for for agent-based
simulations, there is a need to identify the trip purposes, in order to construct the agents daily
plans. To this extent, we begin by looking at the existing activity inference methods for mobile
phone data, followed by studies who have built a mobile phone data driven agent simulation.

4.3.1   Activity  Inference   Methods  

 
 
In traditional survey data, activity purposes are revealed by individuals who answer the travel
diaries; whereas in the mobile phone data, activity types are latent. Furthermore, none of the data
sources (travel survey or mobile phone data) pinpoint the exact locations of the trip destinations,
but areas surrounding these precise locations. Generally, we can find in the literature two
different approaches for Activity Inference — time-frequency rules, and probabilistic models.

Activity   inference   by   time-­frequency   rules   One of the direct ways to infer contextual
information such as location’s function or trip purpose, is by time-frequency rules. Several
authors (Toole et al., 2015; Alexander et al., 2015; Jiang et al., 2015) have improved the general
idea shown in Wang et al. (2012) and Iqbal et al. (2014) in using both visit frequencies and
temporal data to identify work, home, and other locations. A user home location is defined as the
stay point most frequently observed between 8 pm and 7 am during weekdays and weekends.
Whereas, work location is defined as the stay point, other than home, that users visits the most
between 7 am and 8 pm on weekdays. Since some individuals do not work, the work location is
left blank if the candidate location is not visited more than once per week or if the location is less
than 500 m away from the home location (to avoid work identification through signal noise
rather than a distinct location). Another variation (Alexander et al., 2015) is that work location is
identified as the stay to which the user travels the maximum total distance from home location,
to identify evening and night shift jobs. All remaining non-home or work stay points are
designated as other.

23
Transport Modelling in the Age of Big Data June 2016

Activity  inference  by  probabilistic   modelling   Time-frequency rules to infer activity context
are a straightforward method, but may not be useful for certain groups in the population. More
robust inference methods can be applied using probabilistic models, in which other datasources
can be naturally included (semantic-enriched geographic data, POIs, etc.). Thus, we can reference
the spatial information and characteristics of destinations to build probabilistic models to infer
activity types at different destinations in space and time.

Jiang et al. (2013) formulate the target probability as the probability that an individual performs
a certain activity at a certain time depending (conditional) to her/his destination information
and her/his extracted unlabelled activity chain (i.e. motif, (Schneider et al., 2013)). Once the
problem has been reformulated as a target probability, the next step is to choose a probabilistic
inference model suitable to handle the dependencies on the random variables and to calibrate the
parameters of the model. For the latter problem, two approaches exist based on the availability
of training data: — Supervised learning, and Unsupervised learning.

Activity  Classification   Activity classification is done through supervised learning algorithms

and it requires a labelled training set. It addresses the problem of identifying the activity type
for each stay, given the extracted stay sequences and labelled travel survey data and requires to
predefine the categories that the classification algorithm will be trained to detect. For instance,
Yang et al. (2015) work with 5 different activity categories: home, work, leisure, shopping, and
other. For their feature vector they use activity start times, activity duration and the location.
They employed a Bayesian classification algorithm which recognises the transition probability
between activities from the labelled training data. Given that home and work have more distinct
start time, duration and location distribution, they performed a stepwise classification, firstly for
the work and home labels, and secondly to distinguish between leisure and shopping using the
transition probabilities.

Activity   Clustering   When there is no available labelled training data, we can still reveal the
spatial-temporal structure of activities by performing unsupervised clustering. Widhalm et al.
(2015) and Yang et al. (2015) clustered the stays into meaningful categories based on stay start
time, duration, sequence of stay locations, and a vector of land use shares. To model the
dependencies between the explanatory variables and construct the joint distribution they used a
Relational Markov Network (based on the work of (Liao et al., 2005)), which is an extension of
undirected graphical models known as Markov Random Fields. To compute the joint posterior
distribution they used Rejection Sampling, which is a technique that takes samples from the
posterior distribution. Finally, to train the network in an unsupervised way, they used Expectation
Maximisation (EM) algorithm and applied this methodology for case studies in Boston and
Vienna.
24
 Transport Modelling in the Age of Big Data June 2016

The resulting clusters reflect trip chains and activity scheduling patterns that agree well with
data obtained from traditional surveys. Moreover, the comparison between both cities showed
similarity in their clusters. Still, several future improvements are suggested. Firstly, to study the
relationship between the automatically discovered activity clusters and conventional activity
types used in traditional surveys more closely. Secondly, to examine and include the interaction
between land use and travel behaviour (e.g. density, regional accessibility, roadway connectivity,
etc.) in the modelling, for instance how the inclusion of point-of-interest (POI) databases allows
to further improve the accuracy of the results. And thirdly, the inclusion of the results in a
simulation model (i.e. agent-based model) to compare the resulting traffic flows to actual traffic
measurements as a validation step.

4.3.2   Mode   Inference  

 
 
Transportation mode inference from ubiquitous computing devices is a common challenge in the
literature. However, the majority of the methods proposed are based on mobile phone sensors,
such as GPS, accelerometers, and gyroscopes, in which fine-grained sampling is available. On the
light of getting large-scale observations for transport planning purposes in urban areas, a broader
classification can still be made based only on CDRs. These methods infer the transportation
mode by estimating the mobile phone’s speed and associating it with a transport mode. For
instance, Wang, Calabrese, Lorenzo and Ratti (2010) used the information on a trip’s origin and
destination, as well as its travel time to classify the travel mode in three group: car, public transit,
and walking. Firstly, they filtered their dataset to include only trips with distances more than 3
km, and users with update location frequency of more than 1 per hour. They followed by
grouping trips according to their origins and destinations, and performed k-means clustering to
differentiate between the modes. Finally, the results were validated against Google Maps travel
times information.

4.3.3   Synthetic  population  from  Call  Data   Records  

 
 
As we have seen on the lines of this chapter, many studies have been made to mine individual
travel behaviour from mobile phone data for the purpose of having more frequently-updated and
opportunistic collected data to enhance four-step submodels (e.g. trip generation, distribution
and travel mode), and creating dynamic OD matrices. However, one of the challenges still
remaining consists of using mobile phone data in the context of activity-based transport modelling
frameworks and agent-based simulations.

25
 Transport Modelling in the Age of Big Data June 2016

The work of Zilske and Nagel (2014) represents a pilot study of such an approach. They seek to
replace travel diaries with sets of CDRs as input data for agent-oriented traffic simulation. For
this purpose, they generated synthetic CDR data from a MATSim simulation using a plug-in that
introduces cell coverage and a mobile phone usage model for the agents. CDR information is
generated on arrival at or departure from activity locations. After CDRs are generated, they
identified every observed person with a MATSim agent and converted every call information
into an activity. The main limitation on the study is the simplification of the mobile phone data
generation process. The full-spectre of mobile phone data is not represented (e.g. cell handovers,
automatic location updates, mobile internet usage), plus CDR information was not generated for
instances outside the activity locations (no mobile phone activity on-the-go). These resulted in an
underrepresentation of the traffic simulation in MATSim.

However, in their next paper on this topic, Zilske and Nagel (2015) propose to mitigate the
underrepresentation and reduce the spatio-temporal uncertainty by fusing the CDR dataset with
traffic counts. They computed an expansion factor of the population to compensate for the
underestimate demand and matched it to traffic counts. From each trace they created several
agents and used the expanded population as a buffer (with the introduction of a stay-at-home plan
probability) to steer the demand towards matching the known link volume counts. Additionally,
in order that the modelled travel demand matches with the traffic counts, they introduced a
parameter to alter probability by which a particular activity plan is chosen. Intuitively, the offset
was a calculation based on how much a specific choice of a plan contributes to the whole traffic
system fitting to the traffic counts.

4.3.4   Mobile  phone  data  driven   MATSim  

 
 
The latest work in building a fully mobile phone data driven agent-based simulation is the
Smartbay area project which is presented in a paper by Pozdnoukhov (2015). They introduced to
the MATSim environment anonymised CDRs recorded at the spatial resolution of the deployed
mobile phone towers or antennas. Each individual’s travel behaviour is modelled using a hidden
semi-Markov model (HSMM), assuming that a user’s (hidden) activity is influenced by temporal
factors (i.e. day of the week and hours of the day), the type of the previous activity, and observed
factors related to the current activity. After having inferred activity patterns, daily plans for the
virtual population are generated and a MATSim traffic simulation scenario is run. It is important
to note that the simulation output preserves user anonymisation since CDR trajectories are not
involved in the micro-simulation. The agent-based simulation computes the route selection for
each user based on the individual improvement of the utility function and not directly from
CDR traces, avoiding the possibility to recover one specific mobile phone customer from the
output of the simulation.

26
 Transport Modelling in the Age of Big Data June 2016

5   Supplementary  Datasets:  POIs  and   GPS  

 
 
5.1   POIs  
 

Although public transport smart card and mobile phone data have wide coverage across a city’s
population, the lack of purpose for the identified trips needs to be inferred in order to be useful
for activity-based modelling approaches. Temporal features such as activity start time and
duration may be enough to determine primary locations (i.e. work, home) for the majority of the
population. However, for secondary locations in which individuals perform a range of activities
from dining, shopping, and different types of leisure activities, enrichment of the spatial feature
space is needed from the establishments surrounding a detected stop area from the mobility
traces. Such information can be derived from Points of Interest datasets.

In terms of supplementing mobile phone data with POI information, Noulas and Mascolo (2013)
compute the most popular activity within an area delimited by cellular antennas, using CDRs and
Foursquare check ins and places data. They built a set of features to exploit semantic annotations
of Foursquare data and test different supervised learning classifiers for the following classes: Arts
and Entertainment, College and Education, Food, Work, Nightlife, Parks and Outdoors, Shops,
Travel spots. Results show better accuracy for Nightlife, Arts and Entertainment, followed by
Shops and Parks and Outdoors, whilst, the classifiers did not show good results for College and
Education. These results encourage the use of POI as a complement on helping with secondary
activities inference, in which, temporal features are not enough.

A similar work was done by Phithakkitnukoon, Horanont, Di Lorenzo, Shibasaki and Ratti
(2010) for which they used POIs extracted from Yahoo maps API. They cluster four different
types of activities (eating, shopping, entertainment, and recreational) using k-means algorithm,
and estimate the most probable activity in each cluster using Bayes theorem. Using the most
probable activity in a region and work place identification through temporal features (Calabrese
et al., 2011), they infer a daily activity pattern for each of the users. As a result, they found a
strong correlation in daily activity patterns within groups of people who share a common work
area’s profile.

It is important to note that both studies focus on identifying the most probable/popular activity in
a delimited area, and not the most probable activity given a daily individual tour behaviour. For
the latter, it is important to include individual and temporal variables in the activity recognition
process such as the previously visited places and the starting time and duration of the activities,
so that the region-based approach can be changed to an individual-based activity inference
process. While approaches to do so with mobile phones and fare collection smart cards are
limited and have been presented in previous chapter, in the next section we will present the
extensive efforts done to annotate individual trip purposes using GPS traces.
27
 Transport Modelling in the Age of Big Data June 2016

5.2   GPS  
 
GPS consist of a spatio-temporal high-resolution track (i.e. usually one reading per second,
depending on configuration and application purpose). The importance of GPS technology or
mobility applications is growing steadily: it is widely used for location-based smart phone
application, to monitor public transport operations (GPS-enabled buses and taxis), and in the
automobile industry as core technology for various applications such as route guidance but also
is applied for toll collection and travel diary surveys.

However, in this chapter, we focus on applications that use GPS data to derive information with
a semantic meaning (i.e. trip purpose). Depending on their methods, the following taxonomy
according to Huang, Li and Yue (2010) is presented.

5.2.1   Distance-­based  
 
 
The basic idea is to assign the closest POIs to the raw trajectory’s clusters. Commonly, a
minimisation of the Euclidean distance is used between the location of nearby POIs and the
identified GPS stay point (Bohte and Maat, 2009); Xie, Deng and Zhou (2009) present a more
elaborated procedures for which they construct a Voronoi diagram using POIs as Voronoi sites
and then select the POIs closest to the polyline geometry of the trajectory.

Although distance-based methods are easy to implement, they are only suitable for when there
exists a high-accurate GPS trace and therefore are not ideally suited for application to mobile
phone and smart card data.

5.2.2   Attractiveness-­based  
 
 
Different to the distance-based methods, attractiveness-based approached are designed to also
include POI-related information to assign a particular POI given a GPS-point cluster. Huang et al.
(2010) measured the spatio-temporal POI attractiveness based on statical factors (e.g. size of
POI, popularity, and category) and a dynamic function for the attractiveness variation along the
hours of the day. From this information they constructed an attractiveness prism, and selected
the POI for which the mobility trajectory intersects the prism. Furletti, Cintia, Renso and
Spinsanti (2013) present a similar approach, but apply a gravitational model to identify the most
probable activity from a list of ranked POIs.

28
 Transport Modelling in the Age of Big Data June 2016

One of the advantages of the attractiveness-based approach is that it is also suitable for other
types of mobility traces to a certain extent. The CDR + POI examples covered in the previous
section (Noulas and Mascolo, 2013; Phithakkitnukoon et al., 2010) can be fitted in this category,
for the reason that they seek the most attractive/popular activity in an area.

5.2.3   Probabilistic-­based  to  handle  uncertainty  

 
 
Probabilistic-based methods are ideally suitable if certain GPS information is lacking or uncertain.
Typical use cases are when GPS traces are only available for a vehicle a person used but do not
cover walking legs leading to the actual destination or when surrounding high-rise buildings
decrease the signal accuracy. Moreover, probabilistic approaches can handle uncertainty and
capture interdependencies between explanatory variables, which make them a promising
alternative to exploit different streams of information along with low resolution human mobility
sensors. For these reasons, we present in the following an overview of such approaches in a
dedicated section.

5.3   Probabilistic  Graphical  Models  for  activity  inference  

 
 
Probabilistic Graphical Models (PGM) as described in Koller and Friedman (2009) are graphical
representations that efficiently encode and manipulate probability distributions over high-
dimensional spaces. Variables are represented by nodes, and the probabilistic dependency (i.e.
causality) is represented by edges that connect two variables. These graph models can be
regarded as a compact or factor representation of a set of independences that hold in the specific
distribution.

Learning   The structure of a graphical model can be learned from data automatically or pre-
defined by human knowledge. Graphical models usually contain hidden variables to be inferred
(i.e. activity / trip purpose). The learning process of graphical models is to estimate the
probabilistic dependency between different variables given the observed data. Expectation and
Maximisation (EM) algorithms are commonly used methods.

Inference   The inference process is to predict the status of hidden variables, given the values of
observed variables and learned parameters. The inference algorithms can include deterministic
approaches, such as variational methods, and stochastic algorithms like Gibbs Sampling.

In the following, we present representative examples for both generative and discriminative
classification graphical models.
29
 Transport Modelling in the Age of Big Data June 2016

5.3.1   Generative  models:    Hidden  Markov   Models  

 
 
Generative models explicitly model the dependencies between the observations and the class
labels through the joint probability distribution p(x,  y). Hidden Markov Models (HMM) are
generative models represented mainly by an entire observation sequence and a hidden high-level
state sequence connected through a state transitions probability matrix, a matrix representing the
conditional probability between states and observations, and an initial probability distribution.

Different extensions of HMMs have been developed to compute activity likelihoods and
probabilistic estimates of the purpose behind the stop. For instance, Liao et al. (2007b) proposed
a hierarchical HMM trained in an unsupervised manner using Expectation Maximisation (EM)
to learn the parameters of the models, and Rao-Blackwellised particle filter for the inference task.
However, the model still needed to be expanded in order to include information about time of day
and day of the week. Such modification was included in Duong et al. (2005). They introduce
a Switching Hidden Semi-Markov Model (S-HSMM) to exploit both the inherent hierarchical
organization of the activities and their typical durations.

More recently, Yan, Chakraborty, Parent, Spaccapietra and Aberer (2011) developed a framework
that enriches trajectories with any kind of semantic data provided by POI datasets based on a
Hidden Markov Model. To define the initial probabilities they used the percentage of POI
samples belonging to each category. Then they use information on transitions between regions
(annotated with land use information) to construct the state transition. Finally, to infer the hidden
states they maximised the likelihood of the HMM. One of their main contributions is the effective
use of both land use information with POIs to infer the activity purpose. Additionally, Baratchi,
Meratnia, Havinga, Skidmore and Toxopeus (2014) propose an extension of the hierarchical
HSMM that captures spatio-temporal associations in the locational history in both stay-points
and trips connecting the stops.

5.3.2   Discriminative  models:  Conditional  Random  Fields  and  Relational  Markov  

Networks  
 

In contrast with Generative models, Discriminative models avoid making independence

assumptions among the observations, instead they model directly the discriminative boundary
between the different class labels, namely, the model learns the conditional probability
distribution p(y| x). Thus, one of the advantages is that all sorts of rich overlapping features can
be incorporate without violating any independence assumption. (Sutton and McCallum, 2006).  

30
Transport Modelling in the Age of Big Data June 2016
Conditional Random Fields (CRFs) are an example of discriminative models suitable for
classification tasks with complex and overlapped attributes or observations. Liao et al. (2007a)
show a holistic approach using hierarchical CRFs to extract places and activities from GPS
traces. The main objective of their work is to segment a user’s day into everyday activities such
as working, visiting, or travel and to recognise and label significant places such as workplace,
friend’s house, or bus stop. To determine activities, the model relies on temporal features, such
as duration or time of day, and geographic information such as locations of restaurants, stores,
and bus stops. They use maximum pseudo-likelihood estimation to learn the parameters of the
model, and belief propagation for the inference task.

Relational Markov Networks (RMN) are extensions of CRFs that provide a relational language
for describing clique structures and enforcing parameter sharing at the template level. Liao et
al. (2005) trained a RMN for labelling the following activities: at home, at work, shopping,
dinning out, visiting, and other. They incorporate global features (e.g. number of home
locations), temporal information (e.g. duration, time of day), and spatial information (POIs) in
the clique templates. For the inference and learning task they develop a technique based on the
Markov-Chain Monte-Carlo algorithm in a supervised manner. They showed that it is possible
to learn the parameters of a complex model using less data by using priors extracted from other
people’s data.

31
 Transport Modelling in the Age of Big Data June 2016

6   Conclusion  
 
 
6.1   Data  driven  Agent  Based  Modelling  for  Transport  Planning  
 
 
Traditional data sources for transport forecasting, i.e. household travel surveys, are of undeniable
value. They not only cover detailed data on individual and household mobility patterns but also
include relevant information on travel modes and purposes. Yet, stand-alone they are not able to
exploit the full benefits of the agent-based transport modelling paradigm. Two main limitations
can be identified. Firstly, they represent only a small sample of the population (normally around
1%). Secondly, they are usually only updated every five to ten years.4

Opportunistic human mobility sensors tackle these drawbacks and become a promising path to
continue developing agent based models for transport planning. The tradeoff of using such
opportunistic widely-collected information is its raw nature. An additional analytic effort has to
be done to identify trips and trip purposes so that they can be integrated in the agent-based
simulations. Thus, the key challenge is the development of robust algorithms that can extract
daily individual schedules from sparse mobility traces. Specifically, as mentioned in Jiang et al.
(2013), developing effective techniques to link the association rules of semantic land use and
POI information of the diverse areas that individuals visit is an open challenge for estimating the
activity types that individuals engage.

6.2   Model  transferability  from  GPS  to  CDRs  

 
 
One of the directions encountered, when using sparse CDRs to extract activities, is the adaptation
of approaches that originally were developed to be applied with GPS data. For instance, Widhalm
et al. (2015) adopted for CDRs the Relational Markov Networks used in Liao et al. (2005)
originally for GPS traces; or Yuan et al. (2013) and Pozdnoukhov (2015) who adapted Conditional
Random Fields for smart card data and Hidden Markov Models for CDRs, respectively. One of
the reasons is that GPS traces have been the subject of a wider number of studies regarding
activity identification. Hence, one important research question is if those models are suitable
for lower resolution mobility traces, such as the ones provide by CDR and smart card data.
Besides the discrepancy in the level of granularity of the traces, GPS-based studies usually have
a controlled sample with activity labels. This is used to train the parameters of the model. For

4
However, a few authorities have started with continuous surveys also using smart phones to lower the response
burden and increase data quality, especially with regards to capturing activities that only last over a short
duration.

32
 Transport Modelling in the Age of Big Data June 2016

CDRs it is likely that such training sample would not be easy to obtain. One of the possible
options is to use information from travel surveys and design a feature-space in which CDRs
and travel survey information can work along in a transfer-learning paradigm. Finally, another
important issue to note, specifically, about inference models designed from GPS traces, is that
usually they are trained and validated for small samples of the population (e.g. 4 persons in Liao
et al. (2007a)). This certainly rises questions about their performance when scalable to the city
size, in which a wider set of behavioural and mobility patterns from individuals might invalidate
the results achieved by the models.

6.3   Unlocking  the  knowledge  of  different  datasets  

 
 
In order to be able to extract mobility and activity behaviours from human mobility sensors, and
obtain similar quality to travel surveys, we need to combine the information across available
datasets. For instance, if the interest is in finding the transport mode from CDRs, a viable option
is to leverage public transit smart card data and available GPS traces from taxi services in a
probabilistic trajectory matching approach. For the case of activity detection, the inclusion of
POIs to help the inference models from human mobility sensors is a path that has not being
explored yet.

As mentioned in Calabrese et al. (2015), there are some challenges in comparing different
datasets, even if they are related. The main one being the different collection periods and
different spatial units. For instance, census data is usually available at the block level, while
mobile phone data relates to individual cell towers. However, we also see the main advantage of
using different human mobility sensors and supplementary datasets such as travel diary data to
complement underlying the importance of data fusion approaches.

From the literature review on the use of smart card and mobile phone data in transport modelling,
we have seen mainly the use of census or household travel surveys as a mean to validate or to
build expansion factors from the algorithms proposed (Alexander et al., 2015), or approaches
that simply concatenate the features in a classification algorithm (Noulas and Mascolo, 2013;
Chakirov and Erath, 2012). In the Big Data era, however, the aspiration will be to unlock the
power of knowledge from multiple disparate, but potentially connected datasets (Zheng, 2015).
Therefore, we expect the most promising approaches will be applications will stem from data
science domains such as machine learning and data mining.

33
 Transport Modelling in the Age of Big Data June 2016

Cross-­Domain  Data  Fusion    Zheng (2015) presents a good survey that analyses, classifies and
exemplifies methodologies for Cross-Domain Data Fusion. From a transport planning
perspective, the promising methods for activity and transport inference are the probabilistic
semantic-meaning data fusion methods, namely, the Probabilistic Graphical Models reviewed in
the previous chapter. As mentioned before, these models are able to capture the dependencies
and correlations between the features in order to produce better estimates. They also constitute
an essential tool to reason coherently from limited and noisy observations (Koller and Friedman,
2009).

6.4   Data  Privacy  

 
While travel diary data and census information are generally surveyed by government agencies
with consent of the relevant data protection authority, data privacy is an important matter due
to the pervasiveness and level of detail how both smart card and mobile phone data document
an individual’s travel patterns. Even if CDRs are anonymised (smart card data normally is not
linked to a particular person by default), individual mobility patterns are pseudo-identifiers. For
instance, de Montjoye, Hidalgo, Verleysen and Blondel (2013) showed that even with the spatial
resolution given by the mobile phone antennas, four spatio-temporal points are enough to
uniquely identify 95% of the individuals. Efforts have been taken to obfuscate the location in a
way not to be able to re-identify a user and still being able to extract useful mobility patterns.
These algorithmic efforts to preserve privacy are currently lead by the emerging Differential
Privacy. Generally, we expect that addressing those concerns in convincing manner will be
crucial towards the development and practical adoption of data-driven, agent-based simulation
for transport planning.

Differential  Privacy    Differential Privacy (DP) is a mathematical requirement on the results of

interaction with data (Mir et al., 2013). By adding controlled noise, DP formalises the idea that
results of a query should be almost the same whether or not an individual is in the database. DP
hides the participation of a user in a database by choosing a budget parameter that represents the
trade-off between the level of privacy and precision. Andrés et al. (2013) extended the concept
of DP for the protection of location data. Although DP has been proved to be effective with
certain location-based services (Andrés et al., 2013), and with aggregate location information
(Mir et al., 2013), when applied to individual mobility traces it seems that a trade-off between
privacy and precision has not been able to obtain with state-of-the-art techniques (Primault et al.,
2014; Hu et al., 2015).

34
Transport Modelling in the Age of Big Data June 2016

New  Models  of  Data  Ownership    Location obfuscation might not be the viable mechanism for
sharing individual mobility traces. Other suggestions, aside from the algorithmic perspective, are
related to changing the current data ownership paradigm. A popular idea is giving people
ownership of their data (Pentland, 2009). With the creation of data vaults (Mun et al., 2010)
or by means of a data trust (Lawrence, 2016), each individual would have the right to dispose
or distribute their personal information. However, a short-term implementation doesn’t seem
plausible since there exist conflict of interests with big data-driven companies. Still, a transparent
use of personal information is certainly a relevant way to facilitate access for researchers to
both develop better anonymisation methodologies and showcase the societal benefits of using
anonymised mobility data.

35
 Transport Modelling in the Age of Big Data June 2016

7   Summary  and  Research  Agenda  

 
 
7.1   Summary  
 
 
In the first part of the paper, we provide a primer on the latest advances in transport demand
modelling. We acknowledge that activity-based models and agent-based simulations tie in well
with new Big Data sources documenting human mobility since both directly stem from the
concept of individual travel patterns rather then aggregate traffic flows. In order to fully exploit
the capabilities of agent-based simulations, there is tremendous potential to not only use
conventional data inputs (e.g. travel surveys, population census), but also include new
opportunistically collected mobility traces, namely, public transit smart card and mobile phone
data, which can document travel behaviour at an unprecedented scale and level of detail. However,
an additional analytic effort has to be done to identify trips and trip purposes so that they can be
integrated into activity-based travel demand framework and to be used to its full potential in
agent-based simulations.

In the second part of the paper, we present a literature review on the methodologies needed to
extract mobility behaviour from such Big Data sources. From trip identification to activity
inference, and their application for transport demand models, we review the efforts in a step-by-
step manner both for the public transport smart card and mobile phone data. We also cover the
relevance of other datasets such as POIs to infer trip purposes and also document how GPS-based
data collection and Probabilistic Graphical Models can be included for model refinement.

Finally, we discuss the findings of the literature review and also identify a set of future challenges,
in particular with regards to data privacy implications.

7.2   Research  Agenda  

 
 
The biggest potential for future research we identify on the application of big data sources (e.g.
smart card and mobile phone data) in combination with traditional data sources (e.g. household
travel surveys) to inform activity-based models and agent-based simulations for transport
planning. To this end, the following challenges need to be added to the research agenda.

36
Transport Modelling in the Age of Big Data June 2016

1.  Further exploration of probabilistic approaches that can handle the uncertainty of Big data
mobility traces in the modelling process and showcasing the relevance of such approaches for
instance through validation with independent data sets such as loop-detector data.

2.  Integration of different available datasets in a data-fusion scheme. For instance, smart card
data and CDRs for mode inference or POI datasets and semantic information in social networks
to inform in the activity inference process of CDR traces.

3.  Exploration on Transfer Learning approaches to cover up the lack of training samples when
using opportunistic collected data sources.

4.  Seek new ways to validate the results for every step of the data mining pipeline.

5.   Further exploration of preprocessing techniques for sparse and noisy mobility traces. For
instance, demonstrate the effectiveness of non-linear Kalman filters, or more complex probabilistic
filters (e.g. Gaussian processes).

6.  Measure and guarantee privacy in the output of agent-based simulations.

In any case, we also acknowledge that sophisticated modelling knowledge has developed in the
domain of transport planning and therefore we strongly advise that still domain expert knowledge
should build the fundament when applying data driven approaches in transport planning. These
new challenges call for a multidisciplinary collaboration between transport modellers and data
scientists

37
 Transport Modelling in the Age of Big Data June 2016

8   Acknowledgements  
 
We would like to thank Seungjae Lee for organising the Seoul Metropolitan Forum 2016 at the
University of Seoul where an earlier version of this paper was presented. This research has been
conducted at the Singapore-ETH Centre for Global Environmental Sustainability (SEC), co-
funded by the Singapore National Research Foundation (NRF) and ETH Zurich.“

9   References  
 
Adnan, M., F. C. Pereira, C. M. Lima Azevedo, K. Basak, M. Lovric, S. Raveau, Y. Zhu,
J. Ferreira, C. Zegras and M. E. Ben-Akiva (2016) SimMobility: A Multi-scale Integrated
Agent-Based Simulation Platform, paper presented at the Transportation Research Board 95th
Annual Meeting.

Alexander, L., S. Jiang, M. Murga and M. C. González (2015) Origin–destination trips by

purpose and time of day inferred from mobile phone data, Transportation Research Part C:
Emerging Technologies, 58, 240–250.

Andrés, M. E., N. E. Bordenabe, K. Chatzikokolakis and C. Palamidessi (2013) Geo-

indistinguishability: Differential privacy for location-based systems, paper presented at the
Proceedings of the 2013 ACM SIGSAC conference on Computer & communications security,
901–914.

Balmer, M., K. Axhausen and K. Nagel (2006) Agent-based demand-modeling framework for
large-scale microsimulations, Transportation Research Record: Journal of the Transportation
Research Board, (1985) 125–134.

Balmer, M., M. Rieser, K. Meister, D. Charypar, N. Lefebvre, K. Nagel and K. Axhausen

(2009) MATSim-T: Architecture and simulation times, Multi-agent systems for traffic and
transportation engineering, 57–78.

Baratchi, M., N. Meratnia, P. J. M. Havinga, A. K. Skidmore and B. A. K. G. Toxopeus (2014) A

Hierarchical Hidden semi-Markov Model for Modeling Mobility Data, paper presented at the
Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous
Computing, UbiComp ’14, 401–412, New York, NY, USA, ISBN 978-1-4503-2968-2.

Barry, J., R. Newhouser, A. Rahbee and S. Sayeda (2002) Origin and Destination Estimation in
New York City with Automated Fare System Data, Transportation Research Record: Journal
of the Transportation Research Board, 1817, 183–187, January 2002, ISSN 0361-1981.

38
Transport Modelling in the Age of Big Data June 2016

Bohte, W. and K. Maat (2009) Deriving and validating trip purposes and travel modes for multi-
day GPS-based travel surveys: a large-scale application in the Netherlands, Transportation
Research Part C: Emerging Technologies, 17 (3) 285–297.

Bouman, P. (2012) Recognizing demand patterns from smart card data for agent-based micro-
simulation of public transport, Ph.D. Thesis, Department of Decision and Information Sciences,
Erasmus University Rotterdam, The Netherlands.

Bouman, P., E. Van der Hurk, L. Kroon, T. Li and P. Vervest (2013) Detecting activity patterns
from smart card Data, paper presented at the BNAIC 2013: Proceedings of the 25th Benelux
Conference on Artificial Intelligence, Delft, The Netherlands, November 7-8, 2013.

Caceres, N., J. P. Wideberg and F. G. Benitez (2007) Deriving origin destination data from a
mobile phone network, IET Intelligent Transport Systems, 1 (1) 15–26, March 2007, ISSN
1751-956X.

Calabrese, F., L. Ferrari and V. D. Blondel (2015) Urban sensing using mobile phone network
data: a survey of research, ACM Computing Surveys (CSUR), 47 (2) 25.

Calabrese, F., G. D. Lorenzo, L. Liu and C. Ratti (2011) Estimating Origin-Destination Flows
Using Mobile Phone Location Data, IEEE Pervasive Computing, 10 (4) 36–44, April 2011,
ISSN 1536-1268.

Castiglione, J., M. Bradley and J. Gliebe (2015) Activity-based travel demand models: a primer,
Transportation Research Board, Washington, DC, ISBN 978-0-309-27399-2.

Chakirov, A. and A. Erath (2012) Activity Identification and Primary Location Modelling based
on Smart Card Payment Data for Public Transport, Toronto, June 2012.

de Dios Ortuzar, J. and L. G. Willumsen (2011) Modelling transport, John Wiley & Sons.

de Montjoye, Y.-A., C. A. Hidalgo, M. Verleysen and V. D. Blondel (2013) Unique in the crowd:
The privacy bounds of human mobility, Scientific reports, 3.

Devillaine, F., M. Munizaga and M. Trépanier (2012) Detection of Activities of Public Transport
Users by Analyzing Smart Card Data, Transportation Research Record: Journal of the
Transportation Research Board, 2276, 48–55, December 2012, ISSN 0361-1981.

Duong, T. V., H. H. Bui, D. Q. Phung and S. Venkatesh (2005) Activity recognition and
abnormality detection with the switching hidden semi-Markov model, paper presented at the
IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2005.
CVPR 2005, vol. 1, 838–845 vol. 1, June 2005.

Eagle, N. and A. S. Pentland (2009) Eigenbehaviors: identifying structure in routine, Behavioral

Ecology and Sociobiology, 63 (7) 1057–1066, April 2009, ISSN 0340-5443, 1432-0762.

39
Transport Modelling in the Age of Big Data June 2016

Eagle, N. and A. (Sandy) Pentland (2006) Reality Mining: Sensing Complex Social Systems,
Personal Ubiquitous Comput., 10 (4) 255–268, March 2006, ISSN 1617-4909.

Ficek, M. and L. Kencl (2012) Inter-call mobility model: A spatio-temporal refinement of call
data records using a Gaussian mixture model, paper presented at the INFOCOM, 2012
Proceedings IEEE, 469–477.

Fourie, P. J. (2014) Reconstructing bus vehicle trajectories from transit smart-card data, Working
paper, 986.

Fourie, P. J., A. Erath, S. A. Ordóñez Medina, A. Chakirov and K. W. Axhausen (2016) Using
smartcard data for agent-based transport simulation: the case of Singapore, in J.-D.
Schmoecker and F. Kurauchi (eds.) Public Transport Planning with Smart Card Data, Taylor
& Francis.

Furletti, B., P. Cintia, C. Renso and L. Spinsanti (2013) Inferring human activities from gps
tracks, paper presented at the Proceedings of the 2nd ACM SIGKDD International Workshop
on Urban Computing, 5.

Goulias, K. G., C. R. Bhat, R. M. Pendyala, Y. Chen, R. Paleti, K. C. Konduri, T. Lei, D. Tang,

S. Y. Youn, G. Huang and others (2012) Simulator of activities, greenhouse emissions,
networks, and travel (SimAGENT) in Southern California, paper presented at the 91st annual
meeting of the Transportation Research Board, Washington, DC.

Horn, C., S. Klampfl, M. Cik and T. Reiter (2014) Detecting Outliers in Cell Phone Data,
Transportation Research Record: Journal of the Transportation Research Board, 2405, 49–56,
July 2014, ISSN 0361-1981.

Horni, A., K. Nagel and K. Axhausen (2016) The Multi-Agent Transport Simulation MATSim,
Ubiquity Press, June 2016, ISBN 978-1-909188-75-4.

Hu, X., M. Yuan, J. Yao, Y. Deng, L. Chen, Q. Yang, H. Guan and J. Zeng (2015) Differential
privacy in telco big data platform, Proceedings of the VLDB Endowment, 8 (12) 1692–1703.

Huang, L., Q. Li and Y. Yue (2010) Activity identification from GPS trajectories using spatial
temporal POIs’ attractiveness, paper presented at the Proceedings of the 2nd ACM SIGSPATIAL
International Workshop on Location Based Social Networks, 27–30.

International Transport Forum (2015) Big Data and Transport - Understanding and assessing
options, Technical Report, OECD, Paris.

Iqbal, M. S., C. F. Choudhury, P. Wang and M. C. González (2014) Development of ori- gin–
destination matrices using mobile phone call data, Transportation Research Part C:
Emerging Technologies, 40, 63–74, March 2014, ISSN 0968-090X.

40
Transport Modelling in the Age of Big Data June 2016

Isaacman, S., R. Becker, R. Cáceres, S. Kobourov, M. Martonosi, J. Rowland and A. Varshavsky

(2011) Identifying Important Places in People’s Lives from Cellular Network Data, in K. Lyons,
J. Hightower and E. M. Huang (eds.) Pervasive Computing, no. 6696 in Lecture Notes in
Computer Science, 133–151, Springer Berlin Heidelberg, June 2011, ISBN 978-3-642-21725-8
978-3-642-21726-5. DOI: 10.1007/978-3-642-21726-5_9.

Jiang, S., J. Ferreira Jr and M. C. González (2015) Activity-Based Human Mobility Patterns
Inferred from Mobile Phone Data: A Case Study of Singapore, paper presented at the Int.
Workshop on Urban Computing.

Jiang, S., G. A. Fiore, Y. Yang, J. Ferreira Jr, E. Frazzoli and M. C. González (2013) A review
of urban computing for mobile phone traces: current methods, challenges and opportunities,
paper presented at the Proceedings of the 2nd ACM SIGKDD international workshop on
Urban Computing, 2.

Koller, D. and N. Friedman (2009) Probabilistic graphical models: principles and techniques,
MIT press.

Lawrence, N. (2016) Data trusts could allay our privacy fears, June 2016,
http://www.theguardian.com/media-network/2016/jun/03/data-
trusts-privacy-fears-feudalism-democracy.

Liao, L., D. Fox and H. Kautz (2005) Location-based Activity Recognition Using Relational
Markov Networks, paper presented at the Proceedings of the 19th International Joint
Conference on Artificial Intelligence, IJCAI’05, 773–778, San Francisco, CA, USA.

Liao, L., D. Fox and H. Kautz (2007a) Extracting Places and Activities from GPS Traces Using
Hierarchical Conditional Random Fields, The International Journal of Robotics Research,
26 (1) 119–134, January 2007, ISSN 0278-3649, 1741-3176.

Liao, L., D. J. Patterson, D. Fox and H. Kautz (2007b) Learning and inferring transportation
routines, Artificial Intelligence, 171 (5–6) 311–331, April 2007, ISSN 0004-3702.

Mir, D. J., S. Isaacman, R. Caceres, M. Martonosi and R. N. Wright (2013) Dp-where:

Differentially private modeling of human mobility, paper presented at the Big Data, 2013
IEEE International Conference on, 580–588.

Mun, M., S. Hao, N. Mishra, K. Shilton, J. Burke, D. Estrin, M. Hansen and R. Govindan (2010)
Personal data vaults: a locus of control for personal data streams, paper presented at the
Proceedings of the 6th International COnference, 17.

Munizaga, M. A. and C. Palma (2012) Estimation of a disaggregate multimodal public transport

Origin–Destination matrix from passive smartcard data from Santiago, Chile, Transportation
Research Part C: Emerging Technologies, 24, 9–18, October 2012, ISSN 0968-090X.

41
Transport Modelling in the Age of Big Data June 2016

Noulas, A. and C. Mascolo (2013) Exploiting foursquare and cellular data to infer user activity
in urban environments, paper presented at the Mobile Data Management (MDM), 2013 IEEE
14th International Conference on, vol. 1, 167–176.

Pentland, A. (2009) Reality Mining of Mobile Communications: Toward A New Deal On Data,
in Social Computing and Behavioral Modeling, 1–1, Springer US, ISBN 978-1-4419-0055-5
978-1-4419-0056-2. DOI: 10.1007/978-1-4419-0056-2_1.

Phithakkitnukoon, S., T. Horanont, G. Di Lorenzo, R. Shibasaki and C. Ratti (2010) Activity-

aware map: Identifying human daily activity pattern using mobile phone data, in Human
Behavior Understanding, 14–25, Springer.

Pozdnoukhov, A. (2015) Activity Based Travel Demand Modelling with Cellular Data, October
2015, http://www.ucconnect.berkeley.edu/workforce-development/
ucconnect-transportation-planning-workshops/role-big-data-
transportation.

Primault, V., S. B. Mokhtar, C. Lauradoux and L. Brunie (2014) Differentially private location
privacy in practice, arXiv preprint arXiv:1410.7744.

Rasouli, S. and H. Timmermans (2013) Activity-based models of travel demand: promises,

progress and prospects, International Journal of Urban Sciences, 0 (0) 1–30, ISSN 1226-5934.

Sarlas, G. and K. W. Axhausen (2015) Localized Speed Prediction with the use of Spatial
Simultaneous Autoregressive Models, paper presented at the Transportation Research Board
94th Annual Meeting.

Schlaich, J., T. Otterstätter and M. Friedrich (2010) Generating trajectories from mobile phone
data, paper presented at the Proceedings of the 89th annual meeting compendium of papers,
transportation research board of the national academies.

Schneider, C. M., V. Belik, T. Couronné, Z. Smoreda and M. C. González (2013) Unravelling

daily human mobility motifs, Journal of The Royal Society Interface, 10 (84) 20130246, July
2013, ISSN 1742-5689, 1742-5662.

Seaborn, C., J. Attanucci and N. Wilson (2009) Analyzing Multimodal Public Transport Journeys
in London with Smart Card Fare Payment Data, Transportation Research Record: Journal of
the Transportation Research Board, 2121, 55–62, December 2009, ISSN 0361-1981.

Smith, L., R. J. Beckman, D. Anson, K. Nagel and M. E. Williams (1995) TRANSIMS:

TRansportation ANalysis and SIMulation System, paper presented at the 5th TRB National
Transportation Planning Methods Applications Conference, Seattle, April 1995.

42
Transport Modelling in the Age of Big Data June 2016

Sun, L., D.-H. Lee, A. Erath and X. Huang (2012) Using Smart Card Data to Extract Passenger’s
Spatio-temporal Density and Train’s Trajectory of MRT System, paper presented at the
Proceedings of the ACM SIGKDD International Workshop on Urban Computing, UrbComp
’12, 142–148, New York, NY, USA, ISBN 978-1-4503-1542-5.

Sun, L., A. Tirachini, K. W. Axhausen, A. Erath and D.-H. Lee (2013) Models of Bus
Boarding/Alighting Dynamics and Dwell Time Variability.

Sutton, C. and A. McCallum (2006) An introduction to conditional random fields for relational
learning, Introduction to statistical relational learning, 93–128.

Toole, J. L., S. Colak, B. Sturt, L. P. Alexander, A. Evsukoff and M. C. González (2015) The path
most traveled: Travel demand estimation using big data resources, Transportation Research
Part C: Emerging Technologies, 58, 162–177.

Trépanier, M., N. Tranchant and R. Chapleau (2007) Individual Trip Destination Estimation in a
Transit Smart Card Automated Fare Collection System, Journal of Intelligent Transportation
Systems, 11 (1) 1–14, April 2007, ISSN 1547-2450.

Wang, H., F. Calabrese, G. D. Lorenzo and C. Ratti (2010) Transportation mode inference from
anonymized and aggregated mobile phone call detail records, paper presented at the 2010
13th International IEEE Conference on Intelligent Transportation Systems (ITSC), 318–323,
September 2010.

Wang, P., T. Hunter, A. M. Bayen, K. Schechtner and M. C. González (2012) Understanding

Road Usage Patterns in Urban Areas, Scientific Reports, 2, December 2012, ISSN 2045-2322.

Widhalm, P., Y. Yang, M. Ulm, S. Athavale and M. C. González (2015) Discovering urban
activity patterns in cell phone data, Transportation, 42 (4) 597–623.

Xie, K., K. Deng and X. Zhou (2009) From trajectories to activities: a spatio-temporal join
approach, paper presented at the Proceedings of the 2009 International Workshop on Location
Based Social Networks, 25–32.

Yan, Z., D. Chakraborty, C. Parent, S. Spaccapietra and K. Aberer (2011) SeMiTri: a framework
for semantic annotation of heterogeneous trajectories, paper presented at the Proceedings of
the 14th international conference on extending database technology, 259–270.

Yang, Y., P. Widhalm, S. Athavale and M. C. González (2015) Mobility Sequence Extraction
and Labeling Using Sparse Cell Phone Data.

Yuan, N. J., Y. Wang, F. Zhang, X. Xie and G. Sun (2013) Reconstructing Individual Mobility
from Smart Card Transactions: A Space Alignment Approach, paper presented at the 2013
IEEE 13th International Conference on Data Mining (ICDM), 877–886, December 2013.

43
Transport Modelling in the Age of Big Data June 2016

Zhang, Y., X. Qin, S. Dong and B. Ran (2010) Daily OD matrix estimation using cellular probe
data, paper presented at the 89th Annual Meeting Transportation Research Board.

Zhao, J., A. Rahbee and N. H. M. Wilson (2007) Estimating a Rail Passenger Trip Origin-
Destination Matrix Using Automatic Data Collection Systems, Computer-Aided Civil and
Infrastructure Engineering, 22 (5) 376–387, July 2007, ISSN 1467-8667.

Zheng, Y. (2015) Methodologies for cross-domain data fusion: an overview, Big Data, IEEE
Transactions on, 1 (1) 16–34.

Zilske, M. and K. Nagel (2014) Studying the Accuracy of Demand Generation from Mobile
Phone Trajectories with Synthetic Data, Procedia Computer Science, 32, 802–807, ISSN
1877-0509.

Zilske, M. and K. Nagel (2015) A Simulation-based Approach for Constructing All-day Travel
Chains from Mobile Phone Data, Procedia Computer Science, 52, 468–475, ISSN 1877-0509.

44