Feel Old Yet?

Updating Mode of
Transportation Distributions from
Travel Surveys using Data Fusion
with Mobile Phone Data
Eduardo Graells-Garrido1,2 , Daniela Opitz3 , Francisco Rowe4 , Jacqueline
Arriagada5
arXiv:2204.09482v3 [[Link]] 30 May 2023

1 Department of Computer Science, University of Chile; 2 National Center for Artificial

Intelligence (CENIA); 3 Data Science Institute, Faculty of Engineering, Universidad del
Desarrollo; 4 Department of Geography and Planning, University of Liverpool;
5 University of Leeds

Abstract
For correspondence:
egraells@[Link] Up-to-date information on different modes of travel to monitor transport traffic and evaluate
rapid urban transport planning interventions is often lacking. Transport systems typically rely on
Data availability: The Telefónica traditional data sources providing outdated mode-of-travel data due to their data latency,
Movistar mobile phone records
have been obtained directly from infrequent data collection and high cost. To address this issue, we propose a method that
the mobile phone operator leverages mobile phone data as a cost-effective and rich source of geospatial information to
through an agreement between
capture current human mobility patterns at unprecedented spatiotemporal resolution. Our
the Data Science Institute from
Universidad del Desarrollo and approach employs mobile phone application usage traces to infer modes of transportation that
Telefónica R&D. This mobile phone are challenging to identify (bikes and ride-hailing/taxi services) based on mobile phone location
operator retains ownership of
data. Using data fusion and matrix factorization techniques, we integrate official data sources
these data and imposes standard
provisions to their sharing and (household surveys and census data) with mobile phone application usage data. This integration
access which guarantee privacy. enables us to reconstruct the official data and create an updated dataset that incorporates
Anonymized datasets are available
from Telefónica R&D Chile for
insights from digital footprint data from application usage. We illustrate our method using a case
researchers who meet the criteria study focused on Santiago, Chile successfully inferring four modes of transportation:
for access to confidential data. The mass-transit (all public transportation), motorised (cars and similar vehicles), active (pedestrian
other datasets used in this study
are publicly available. and cycle trips), and taxi (traditional taxi and ride-hailing services). Our analysis revealed
significant changes in transportation patterns between 2012 and 2020. We quantify a reduction
Acknowledgments: We thank in mass-transit usage across municipalities in Santiago, except where metro/rail lines have been
GPA Movistar for facilitating the
data for this study, and Leo Ferres
more recently introduced, highlighting added resilience to the public transport network of these
for insightful discussion. E. infrastructure enhancements. Additionally, we evidence an overall increase in motorised
Graells-Garrido was partially transport throughout Santiago, revealing persistent challenges in promoting urban sustainable
funded by ANID Fondecyt
#1211490 and National Center for transportation. Findings also point to a rise in the share of taxi usage, and a drop in active
Artificial Intelligence CENIA mobility, suggesting a modal shift towards less sustainable modes of travel. We validate our
FB210017, Basal ANID. D. Opitz
findings comparing our updated estimates with official smart card transaction data. The
was partially funded by ANID
#PAI77190057 and ANID Fondecyt consistency of findings with expert domain knowledge from the literature and historical
de Iniciación #11220799. transport usage trends further support the robustness of our approach.

Competing interests: The author

declare no competing interests.

Graells-Garrido et al. | arXiv | June 1, 2023 | 1–31

 Introduction
Understanding travel behaviour is key to creating more resilient and sustainable urban transport
systems. The global collective contribution of the transport sector to carbon emissions was es-
timated to be over 14% of the global estimate (Allen et al., 2012). As cities expand and absorb
population growth, global warming (Acar and Dincer, 2020), air pollution (Colvile et al., 2001), traf-
fic congestion (Thomson, 1998), and urban sprawl (Brueckner, 2005) exert increasing pressure on
decarbonising transport networks and prioritising more sustainable modes of mobility.
However, cities often lack appropriate data analytics to evaluate and implement transport plan-
ning interventions to achieve these goals. Traditional data sources, such as travel surveys, manual
count studies, censuses and stated preference surveys have conventionally been used to estimate
traffic counts, potential demand for transport and assess the spatial patterns of human mobility
within cities (Zannat and Choudhury, 2019). Yet, these data sources are expensive, infrequent and
suffer from data latency (Rowe, 2023). Thus, while these data from traditional sources may provide
an accurate representation of urban transport networks at a given time, such representations may
be unreliable or outdated a few years into the future, particularly in fast-expanding cities or those
that experience shocks due to social unrest, natural hazards (Rowe, 2022) or pandemics (Green et
al., 2021).
Mobility information obtained from sources such as the Global System for Mobile Communica-
tion (GSM) network can help overcome these deficiencies (Rowe, 2023). Mobile network data, when
they are available, provide a cost-effective source of geospatial information to capture human mo-
bility at unprecedented geographic and temporal granularity (Zannat and Choudhury, 2019). Mo-
bile phone data also offer information to capture travel behaviour in real or near-real time, pro-
viding an opportunity to frequently update information about how people move around cities to
support transport planning (Antoniou et al., 2011). However, they only provide geospatial informa-
tion . They do not deliver data on users’ personal attributes, such as age, gender, and income which
are often associated with different travel behaviours, or information on mode of transportation. In-
formation on modal split is however a key ingredient to developing transportation interventions to
reduce private car usage and promote sustainable transport modes (Rodríguez-Núñez and García-
Palomares, 2014).
To leverage the strengths and address these limitations of traditional and new emerging forms
of data, we develop a generalisable Data Fusion (DF) framework to update area-level estimates
of mode of transport trip counts (defined as mode split). DF involves the integration of multiple
related datasets to render a unified representation that facilitates the identification of relevant
patterns.
In transportation, DF has been used for multiple purposes, such as traffic demand prediction
and forecasting (El Faouzi et al., 2011). Unlike prior work, we use DF to produced updated mode
split estimates from travel survey data by integrating inferred origin-destination mode of trans-
port distributions from more up-to-date mobile phone data from Deep Packet Inspection (DPI),
Extended Detail Records (XDR), and socioeconomic and demographic information from a repre-
sentative household survey and census data. XDR and DPI data are used to infer more up-to-date
origin-destination mode of transport estimates. XDRs provide geospatial locations of users, and
DPI data offer information on application usage of related transport modes to infer these estimates
based on existing travel survey data. Household survey and census data are used to incorporate
data on the socio-economic and demographic composition of the resident population and adjust
biases for distributional biases relating to these attributes in mobile phone data.
Our work contributes to expanding the approach followed by Graells-Garrido et al. (2018b), to
estimate the modal split but using a DF framework by incorporating data: (1) from mobile phone ap-
plication records to infer the use of bike-sharing platforms and taxi services to improve estimates
of bicycle and taxi transport modes; and (2) from socioeconomic and demographic attributes of
the local resident population to integrate information about variation in the usage of different

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 2 of 31
 transport modes.
This paper is structured as follows. The Background section delves deeper into the existing
literature and research related to transport data collection and analysis methods regarding the
modal split. In the Context and Data section, we present the specific context of our study, focusing
on Santiago, Chile, and describe the data sources used, including household surveys, census data,
and mobile phone application usage traces. The Methods section outlines our approach, which
involves leveraging mobile phone location data and employing data fusion and matrix factoriza-
tion techniques to integrate official data sources with digital footprint data. In the Results section,
we present the findings of our analysis, including insights into transportation patterns and shifts
between 2012 and 2020 in Santiago, and a thorough analysis and interpretation of the results, com-
paring them with expert domain knowledge and validation data. The Discussion section explains
the practical and theoretical implications of our work, as well as its limitations and future lines of
research. Finally, the Conclusions section closes the paper with a highlight of our contribution to
knowledge and transportation planning.

Background
Information on transportation modes is critical to urban transport planning. It enables improving
the efficiency of transport systems by identifying suitable options for specific travel needs (e.g.,
long-distance commuting or freight transport) to design multimodal transportation networks (e.g.,
Karlaftis (2004)). Accurate information on modes of transport also enables assessment of trans-
port demand by identifying locations of high demand for specific travel options to inform the ap-
propriate allocation of resources (Karlaftis, 2004). Different modes of transport require different
types of infrastructure and have varying environmental footprint (e.g., Vuchic (2007)). Having cur-
rent information on modes of transport is thus essential to identify critical infrastructure needs
and develop strategies to reduce carbon emissions and promote more sustainable forms of urban
mobility (Vuchic, 2007).
The estimation of transportation modes using traditional statistical methods and machine learn-
ing techniques is a well-established problem in transportation research. Traditionally, estimates of
mode split have been derived from data collected through household travel or travel diary panel
surveys, involving retrospective, self-reporting records of how often people use various modes of
transport (Lee et al., 2020; Molin et al., 2016), or of individuals’ trips for a selected number of days
(Heinen and Chatterjee, 2015; Kuhnimhof et al., 2006). Data collection through these surveys is
however costly and time-consuming posing significant limitations to their scalability to cover large
geographical areas, population samples and complex multimodal systems. In recognition of these
limitations, increasing efforts have been made to leverage on the emergence of digital footprint
data to estimate transportation modes (e.g., Graells-Garrido et al. (2018b), Bachir et al. (2019)).
Two forms of digital footprint data feature prominently in the estimation of travel counts by
transportation modes: (1) mobile network data (including DPI and XDRs), and (2) GPS data. GPS
data have been used to estimate mode of transport use by leveraging on the accuracy of GPS tech-
nology. They provide accurate location information to render individual travel trajectories in real
time by triangulating records on speed, time and location (Bantis and Haworth, 2017; Dabiri and
Heaslip, 2018). However, the availability of individual GPS records is ethically sensitive, geographi-
cally sparse and limited to a small set of users as their collection relies on the use of tracking mobile
phone applications (Rowe et al., 2022). Only a segment of the population uses a given application
or allows tracking (Grantz et al., 2020). Tech companies (e.g., Apple) have included features to
prevent tracking by mobile phone applications and new legislation (e.g., GDPR) has tightened the
use of mobile phones to collect and store personal data. All these issues challenge the use of GPS
records to accurately capture human mobility and hence derive reliable mode split estimates.
In contrast, mobile network data generally cover entire geographical areas (e.g. countries) and
population systems (i.e. users of a network operator). They are also a by-product of an administra-

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 3 of 31
 tive process and thus continuously generated by network operators. As a result, mobile network
data offer more extensive coverage and are more readily available than GPS data. However, mobile
network data are geographically sparse and irregular as their collection depends on the location
and facing position of telecommunication antennas. They do not provide exact information on
location. Users’ location needs to be inferred from their position relative to telecommunication
antennas. This problem represents a challenge to accurately estimate the use of different modes
of transportation at a given location.
A large number of studies have focused on the use of mobile network data to address this chal-
lenge and estimate the use of different modes of transportation(Huang et al., 2019). A large body
of the literature has focused on estimating the use of trains and cars, particularly using data on
intercity trips (Doyle et al., 2011; García et al., 2016; Hui et al., 2017; Schlaich et al., 2010; Smoreda
et al., 2013; Wu et al., 2013). The focus on intercity trips may have been determined by the accuracy
of the estimates that can be produced inferring train and car trips from relatively low spatial and
temporal resolution mobile phone network data. Less work has focused on inferring trips by alter-
native modes of transportation, such as bus, trams, bikes and walk (Graells-Garrido et al., 2018b;
Holleczek et al., 2015; Horn and Kern, 2015; Li et al., 2017; Wu et al., 2013). Alternative modes of
transportation are often aggregated into general groupings, such as public transport versus pri-
vate transport (Horn et al., 2017; Phithakkitnukoon et al., 2017; Qu et al., 2015; Wang et al., 2010),
air versus ground (Hui et al., 2017), moving versus stationary (Calabrese et al., 2011) or rail versus
road (Asgari, 2016; Bachir et al., 2019; Doyle et al., 2011). While useful in specific settings, such
aggregations have more limited utility in the current policy context seeking to develop planning in-
terventions to promote active forms of mobility, identify associated critical infrastructure and thus
reduce carbon emissions.
Previous work seeking to infer mode of transportation from mobile network data has largely
used non-supervised or semi-supervised machine learning algorithms. Notable studies include
Bachir et al. (2019), Graells-Garrido et al. (2018b), Kalatian and Shafahi (2016), Wang et al. (2010).
Wang et al. (2010) and Kalatian and Shafahi (2016) employed non-supervised machine learning
algorithms and mobile phone data to cluster travel times and travel speeds to identify different
modes of transport. Wang et al. (2010) identified two forms of transport (i.e., road and public
transport), while Kalatian and Shafahi (2016) distinguished between walking, public transportation
and private car. Using data from Santiago (Chile), Graells-Garrido et al. (2018b) proposed a method
based on semi-supervised Non-negative Matrix Factorization (NMF) to various transport modes in-
cluding: rail, car, bus, rail & bus and pedestrian modes. Bachir et al. (2019) employed agglomerative
hierarchical clustering to group Voronoi sectors of Paris with similar transport usage and identify
rail, road, and multimodal (rail+road) modes. A novelty of their work was to use census data to
rescale the resulting origin-destination estimates to improve their statistical representation.
While these studies have progressed our capacity to infer mode split patterns from mobile
network data, key challenges remain. For example, Wang et al. (2010) and Kalatian and Shafahi
(2016) only distinguished a selected number of transport modes and are based solely on travel
times or speed travel, which can vary at different traffic stages and show similarities under certain
conditions, particularly in congested cities. The model proposed by Graells-Garrido et al. (2018b)
can infer multiple modes, including road, public transport, rail, pedestrian, and rail+bus, but it
tends to underestimate pedestrian mode results. On the other hand, the model by Bachir et al.
(2019) considers a bimodal partition (rail and road) and does not account for the intermodality of
trips.
We argue that borrowing information from disparate data sources is key to inferring different
modes of transportation. NMF or traditional clustering methods may not be the most suitable ap-
proaches to integrate different data sources as they do not directly take advantage of systematic
relationships identified between attributes across multiple datasets. We argue that data fusion
algorithms can provide a more robust approach to integrate disparate data sources as they are
explicitly designed to use systematic patterns observed across various datasets to fuse them. This

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 4 of 31
 fusion is achieved by quantifying and identifying mathematical patterns in the dataset as if they
were part of a single model. Data fusion algorithms have been predominantly applied in biomed-
ical research (Li et al., 2018) and the modelling of urban lifestyles (González, 2019). Their applica-
tion in urban transportation contexts is less prominent although not absent (El Faouzi et al., 2011).
In our context, they have been used with travel surveys to update modal split estimates (Chang
et al., 2019). However, they have not yet been employed to integrate mobile phone data. We
however recognise that the concept of data integration is not new and has been incorporated in
supervised algorithms of modal split estimation by Huang et al. (2018). They used it to combine
mobile phone signalling data with subway smart card data and taxi GPS data to develop a machine-
learning model that predicts one-hour ahead traffic population counts by area in Shenzhen (China).
We also recognise a variety of applications of supervised machine learning approaches to estimate
modal split based on mobile phone data. These include Chen et al. (2017), Semanjski et al. (2017)
and Breyer et al. (2022). Chen et al. (2017) uses the magnetic signal combined with the accelerom-
eter presented in intelligent devices of different trips labelled with seven modes (stationary, bus,
bike, car, train, light rail and scooter). Semanjski et al. (2017) uses labelled trips collected through
the application Routecoach combined with urban infrastructure obtained from OpenStreetMap
(OSM), to predict the modes of travel: driving, public transport, biking, and walking in the Flemish
Brabant province (Belgium). On the other hand, Breyer et al. (2022) proposes a semi-supervised
model to classify trips between cities from mobile network data. A key limitation of supervised ma-
chine learning approaches is that they require labelled data identifying the modes of travel. Such
a labelled dataset is difficult to generate at scale because they are costly and normally cover a
reduced geographical area.
Thus, significant progress has been made to generate area-level estimates of transportation
modes and to identify various modes of transportation based on travel surveys and mobile phone
data. Most existing work has relied on non-supervised or semi-supervised machine learning al-
gorithms due to the limited availability of labelled data identifying transportation modes. Less
progress has, however, been made in inferring trips undertaken by bike and emerging forms of
shared mobility, such as ride-hailing services. And, while data fusion approaches have displayed
great potential to integrate disparate datasets and identify systematic patterns in order to infer
modal split, they have not been widely adopted to produce modes of transportation estimates.
This paper seeks to propose a data fusion framework to integrate various data sources (list data
sources) to estimate the extent of bike and ride-hailing service usage, along with other public and
motorised forms of mobility.

Context and Data

Our work studies the Great Metropolitan Area of Santiago, Chile. Santiago is a city with almost
8 million inhabitants and an urban surface of 867.75 square kilometres as of 2017, composed
of more than 40 independent municipalities, 35 within the urban area, whereas the rest are on
the periphery. The city has an integrated multimodal system with an almost flat fare between
metro, urban buses, and one rail service, allowing up to three legs within a two-hour window. The
bip! smart card is the only accepted payment method in the public transportation system. The
city also has ubiquitous taxi services and access to several ride-hailing applications, and a public
bike-sharing system. For this work, we congregate all modes of transportation into four categories:
mass-transit, which includes all public transportation that uses the city smart card; motorised, which
refers to cars and similar vehicles; active, which includes pedestrian and cycle trips; and taxi, which
includes traditional taxi and ride-hailing applications.
The last official mode split for Santiago was published in 2015, based on a travel survey from
2012. The survey is outdated considering the different interventions and changes in the city, mak-
ing it difficult to understand current travel behaviour patterns. We emphasise three of these
changes. First, the population composition of the city has significantly changed due to external

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 5 of 31
 and internal migration waves. Second, new transportation infrastructure was built, including two
metro lines and a new train service from 2017 onwards, and new mobile phone applications and
other transportation systems appeared, including ride-hailing apps such as Uber, electric scoot-
ers such as Lime, and shared bike-systems such as Mobike. And third, the social upheaval from
October 2019 caused a series of massive demonstrations and multiple riots throughout the city,
including burned metro stations (Urquieta Ch., 2019), the closing of commercial districts, and an
associated economic crisis (Somma et al., 2021).
To estimate an updated mode split, we integrated the following data sources about Santiago:
mobile phone data from 9—13 March 2020, official data sources from 2020, urban infrastructure
in 2020, and the travel survey collected in 2012. We focus on trips made during morning peak
periods (between 6AM and 9AM) on weekdays. In this way, we try to capture regular activity-based
trips, such as work or study. On the other hand, the morning peak period is the most congested,
and therefore the most interesting period from a planning perspective.

Urban Infrastructure from OpenStreetMap (OSM)

OSM is a global, open geographical database that is freely available and fairly accurate for many
cities (Haklay, 2010; Zhang and Pfoser, 2019). It contains several types of geographical features, in-
cluding urban infrastructure networks. OSM data can be downloaded freely, either from the OSM
organisation itself or in data aggregators such as GeoFabrik. We parsed OSM data from March
2020 in Santiago using the pyrosm tool (Tenkanen and contributors, 2022). Then, we identified
the following urban network information: highways, primary (or main) streets, secondary streets,
tertiary streets, rail and metro networks, cycleways, and pedestrian streets (see Figure 1). These
features are arguably associated with several modes of transportation. We assume that cycle-
ways and pedestrian streets increase the chance of observing active trips in mobile phone records
located near them, whereas highways increase the chance of observing motorised and taxi; we
assume that mass-transit is associated with primary streets and the rail network as well.

Figure 1. Urban infrastructure from OpenStreetMap in March 2020.

Official Data Sources from 2020

We integrated the following official data sources: a socio-economic characterisation survey (2020),
a travel survey (2012), statistics regarding car permits and global metro usage from the National
Institute of Statistics (2020). The travel survey provides an initial estimation of the mode split,

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 6 of 31
 whereas the other sources are used to update the initial estimate, jointly with the mobile phone
data described earlier. In addition, we incorporate origin-destination matrices for mass transit ac-
cording to smart card data for 2012 and 2020 to validate the model. We describe these datasets
in the section Smart Card Data.
The survey for socio-economic characterisation is named CASEN (Ministerio de Desarrollo So-
cial y Familia, 2020). Its 2020 edition is the last one released at the time of reporting this work.
Relevant questions for our work include the income distribution per municipality (in quintiles, see
Figure 2.a for the distribution of the richest quintile), the distribution of professions and occupan-
cies per municipality, and the distribution of migrants with respect to their country of origin per
municipality. These variables are relevant for our methods as Santiago has experienced migration
waves in recent years and a social outburst in 2019. Both have caused changes in work habits,
such as an increase in remote work in areas affected by political protests, and mobility habits, as
migrants have different mobility choices than locals (Rowe and Bell, 2020).
The Santiago Travel Survey (EOD, from its name in Spanish) was collected in 2012 by the Chilean
Ministry of Transport and Telecommunications (SECTRA, 2015). As our mobile phone data were
from labour days within a week in March 2020, we estimated an initial mode split for the munici-
palities under study in a workday in morning commute hours (see Figure 2.b for the mass-transit
municipal share, c.f. the income distribution of Figure 2.a, which is negatively correlated). This esti-
mation aggregates all trips originating at each municipality per mode of transportation, and sums
their survey weights. Following the same procedure, we also estimated the speed range distribu-
tion per mode of transportation (see Figure 3.a), as it can be a relevant feature when discriminating
the mode split in OD flows (Wang et al., 2010); however, these distinctions may be unreliable in
high traffic areas, which is something common at commuting times due to how work areas are
concentrated in Santiago (Suazo-Vecino et al., 2019). Additionally, since the travel survey contains
income information, it is also possible to estimate a mode split per income quintile, that is, the
total sum of survey weights of trips per mode performed by people with a given home income.
There is a high association between motorised transport and higher income levels (see Figure 3.b),
and thus, imputed income according to home municipality could be a relevant signal for mode
inference.

Figure 2. Geographical distributions of: a) the richest household income quintile, from CASEN (2020); b) share
of mass transportation in morning commute time, per municipality, from the Santiago Travel Survey (2012).

The National Institute of Statistics (INE, from its initials in Spanish) publishes yearly informa-

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 7 of 31
 Figure 3. Distribution of speed ranges (a) and income quintiles (b) per mode of transportation (lowest income
is Q01, highest income is Q05). Each matrix is column-normalised.

tion regarding car permits per municipality, smart card transactions at the whole Metro (subway)
network, and population composition in 2012 and 2020 (Instituto Nacional de Estadísticas, 2022).
These statistics inform our method. On the one hand, car permits must be renewed yearly, and,
as such, they are a proxy of how much the motorised modal share has changed: on average, car
permits have increased by 31.76% between 2012 and 2020. On the other hand, the number of
smart card transactions at Metro stations has decreased by 43% from 2012 to 2020, even though
there have been new metro and rail lines since 2012. There are potential explanations for this,
including the effect of the social outburst, where several stations were burned. As such, the official
sources indicate that motorised usage has increased globally, and Metro usage, which is part of
mass-transit, has decreased.
We expect these official sources to aid our proposed model when estimating an updated mode
split. Some of these variables may have importance when estimating the mode split (such as in-
come, car permits, etc.). Other variables provide demographic controls for the model, which may
help to modulate potential biases in the anonymised data (such as the population distributions
from 2012 and 2020).

Mobile Phone Data

We used two mobile phone datasets provided by the telecommunications operator Telefonica Mo-
vistar in Chile: aggregated trajectories from eXtended Detail Records (XDRs), and aggregated traffic
from Deep Packet Inspection (DPI). Both datasets were generated from 1.18M phones active in the
period and area under study. The operator has nearly one-third of the market in the country. In
the urban area of Santiago, it operates 2076 towers distributed in the area of study (see Figure 4.a).
Note that towers are not distributed uniformly in the city. According to the mobile phone opera-
tor, the distribution follows mobile population patterns in order to ensure quality of service. We
studied municipalities with at least 5 towers in the area under study, resulting in 40 municipalities,
with an average of 48.27 towers (see Supplementary Material A1 for a characterization of the tower
distribution). As mentioned in the previous data description, we focused on the morning commute
time (from 6 AM to 9 AM), as it is the most generalisable period that can be studied.
The XDRs dataset is derived from the mobile phone network and is generated through the
polling of device locations (Huang et al., 2019). Each record in the dataset represents a billable
event, including calls, SMS, and Internet downloads. It is important to note that a single record
may encompass multiple events, such as multiple downloads or multiple connections during a
phone call. To determine the most representative tower for these events, a criteria defined by
the company is employed. For calls, this typically corresponds to the originating call tower, while
for Internet downloads, it is often the most frequently used tower within a given set of events.
Furthermore, not all records are generated equally, as the company profiles data usage by its users

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 8 of 31
 and has different billing frequencies depending on how active each customer is. On average, the
records are generated in 15-minute intervals for each device in the Telefónica Movistar network in
Santiago (Graells-Garrido et al., 2016). Each record contains the ID of the representative tower the
device was connected to, a timestamp, and an anonymised device ID. Device IDs are consistent
in the dataset, i.e., two connections with the same ID describe the trajectory of the same device.
We used XDRs to compute trips using a stop-based method (Graells-Garrido et al., 2018b). This
method compares the timestamp and distance between two consecutive events. If the speed of
the transition is above a threshold of 0.5 Km/h, the transition encodes part of a trip. This threshold
discards slow transitions that could be associated with movement inside a venue, for instance, the
device of someone who works in a university will connect to different towers when moving around
the venue. Adjacent parts of a trip are chained to identify its origin and destination towers, as well
as the intermediary towers that serve as waypoints for each trip. In total, we identified 19M trips
in the whole period of study. Of these, 3M were detected during the morning commute period,
between 6 and 9 AM (Figure 4.b shows the total influx minus outflux of trips for each municipality
at that period).

Figure 4. a) Distribution of mobile phone towers (also denoted waypoints) in the area under study. Each dot
is a mobile phone tower. b) Each municipality is coloured according to the difference between inbound and
outbound trips; red municipalities have a positive flux, that is, attracting more trips than those generated by
them, whereas blue municipalities have a negative flux. Red areas concentrate work and wealthy areas in the
city.

Each trip has features, such as date and time, average speed, and waypoints. We observe that
the trip distribution through all days resembles a typical city routine with two main peaks in trip
volume, one in morning commute times and another in afternoon commute times (see Figure 5.a).
The distribution of trip speed is skewed toward lower speeds (see Figure 5.b), which is expected
in an urban setting due to mainly commuting traffic. Note that we analysed trips with speeds
between 5 Km/h, which is just below the average human speed walking (Browning et al., 2006),
and 120 Km/h, which is the legal maximum speed in Chile. Of the 3M trips, only 7.4K had speeds
above 120 Km/h, and 877K had speeds below 5 Km/h. Lastly, waypoints are the intermediary
towers that were part of the trajectory of each trip. These waypoints are important when inferring
modes of transportation, as some waypoints are strongly associated with each mode, such as
towers installed within metro stations or near highways (Graells-Garrido et al., 2018b).
The DPI data contain the number of connections in each tower to the 963 most accessed In-
ternet domains from mobile phones (note that, unlike XDRs, this is not device-level data). These

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 9 of 31
 Figure 5. a) Distribution of trips observed in mobile phone data according to the time of the day (purple line).
Each line marks the start of a period of time defined in the Santiago Travel Survey. In this paper, we focus on
commuting times at morning peak. b) Histogram of speeds in observed trips from XDRs.

domains can be mapped to mobile phone applications or Web services, as company or service
names are encoded in domain addresses. The number of domains was determined after manu-
ally removing spyware or other malicious domains, ad-tracking and other types of analytics per-
formed on apps, removing apps that showed concentration of usage in the city, that is, they are
accessed from a small set of locations (defined as those in the lower 10% of entropy with respect to
the number of requests per tower), and unifying domain names to avoid duplicate app identifiers,
for instance, the domains [Link] and [Link] are unified into [Link].
Thus, whereas XDRs indicates trajectories of devices, DPI indicates digital activities performed by
the people connected to each tower through mobile phone applications (Graells-Garrido et al.,
2022). Some of these applications may aid the identification of previous hard-to-detect modes,
such as pedestrian, non-motorised and taxi (due to ride-hailing applications such as Uber and Cab-
ify). Other apps or activities are important too, as they may be held during transportation (Jain and
Lyons, 2008). Previous work has found that many types of activities or applications are relevant
to identify mobility patterns (Graells-Garrido et al., 2022, 2018a), as such, we analyse a wide spec-
trum of apps, without focusing on transportation apps only. For instance, they may be related to
the mode of transportation of the device owner (Graells-Garrido et al., 2018a). From a total of 1278
non-duplicate apps (or domain names), we kept 963 for analysis after filtering using the entropy
criteria.
As app usage may be related to the mode split by associating waypoints (as seen in XDRs) with
activities (as seen in DPI), we estimated the distribution of mobile phone applications accessed at
each waypoint (tower). This distribution can be expressed as a matrix (see next section for details
of this and other matrices) where each cell contains the total access count to an app reported in
the data at a given tower. We applied the Log-odds ratio with Uninformative Dirichlet Prior formula
to the access counts per application (Monroe et al., 2008). This formula considers the frequency
and variability of access to each application, and determines which waypoints had relevant values
over a Dirichlet prior. Figure 6 displays the association of each tower at commuting times with
example applications after applying this transformation: a) Uber, a ride-hailing app related to taxi;
b) Cabify, another ride-hailing app, related to taxi; c) Waze, a GPS routing, related to motorised and
taxi; d) Transantiago, the official domain of the mass-transit system; e) Niantic Labs, the creator of
augmented reality games such as Pokémon Go, which could be related to active transport (Graells-
Garrido et al., 2017); and e) Spotify, an audio app with a distribution that resembles the metro
network (mass-transit).
Thus, with mobile phone data, there are observations of trips, with their corresponding speeds
and waypoints, and information associated with those waypoints. To the extent of our knowledge,
this combination of XDRs and DPI has not been used in modal split problems before. Then, we

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 10 of 31
 Figure 6. Distribution of mobile application usage (from DPI data) in towers. Each tower is depicted with a
circle with area proportional to the intensity of the application usage. a) Uber (ride-hailing). b) Cabify
(ride-hailing). c) Waze (GPS routing). d) Transantiago (mass-transit). e) Niantic Labs (augmented reality games
such as Pokémon Go). e) Spotify (audio). Black lines represent primary streets in the city, cyan lines represent
the metro network. Urban network data has copyright from OpenStreetMap contributors, used with
permission.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 11 of 31
 expect to obtain an updated modal split by integrating these data with the official data previously
described.

Smart Card Data

Finally, we include an additional dataset from the system ADATRAP (Gschwender et al., 2016) pro-
vided by the Direction of Metropolitan Public Transport (DTPM from its initials in Spanish) for vali-
dation purposes. ADATRAP integrates multiple data sources from the mass transit system in San-
tiago, including smart card data, bus GPS data, vehicle operator data, among others. In Santiago,
passengers must validate their bip! cards when boarding a bus or entering a Metro station, but no
alighting validation is required, thus, smart card data only registers the boarding location of trips.
the methodology relied on usage patterns such as the regularity of commuting trips to estimate
the number of trips made. Hence, a methodology based on recurrent patterns of usage per user
and geometry of distances toward boarding sites is used (Munizaga and Palma, 2012). The ADA-
TRAP methodology also includes fare evasion correction (Munizaga et al., 2020). Trips are reported
at municipality and macro-area levels (a macro-area contains multiple municipalities).

Figure 7. Distribution of trips per municipality in smart card data from ADATRAP.

To ensure a rigorous validation of our results, ADATRAP estimated mass-transit trip counts from
commuting times using smart card data between March 9, 2020 to March 13, 2020, from buses
and metro . Specifically, we focused on trips during the morning peak period (06:00 AM - 09:00
AM). Given that ADATRAP provides trip counts with the corresponding time of day and date, we
calculated the average daily trip count during the commuting times for the specified dates and
study period. In total, the system reported 932K trips in mass transit during the morning peak
period in Santiago for the analysed dates.

Methods
Here we propose a method that processes traditional and digital data sources (see an overview of
the method in Figure 8). The method consists of three main stages:
1) Initial mode split estimation, where official sources and domain > knowledge are used to build
a candidate mode split for all > administrative units.
2) Data fusion using matrix factorization, a procedure that generates > a latent representation of
several datasets represented in an > unified manner. As data we consider official and digital
sources, > with features related to mobility patterns, socio-demographic > characteristics,
and urban infrastructure.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 12 of 31
 3) Updated mode split estimation, where generate an updated mode > split that takes into ac-
count all available information in the > latent representations from the previous point.

Official Data (representative, coarse resolution)

Speed Metro Car Projected Professions/ OD Flows
OD Flows Income Migration
Distribution Transactions Permits Population Occupations
Direction of
Socio-Economic Metropolitan
Travel Survey National Institute of Statistics
Characterization Public Transport
(2012) (2020)
(2020) (2020)
Municipal City/Municipal Municipal
Municipal

Initial Data Fusion Updated

Mode Split using Municipal Mode Split
Estimation Matrix Factorization Estimation

Tower

Mobile Phone Mobile Phone

OpenStreetMap
Billing Records Network Traffic
(2020)
(2020) (2020)
Speed Application
OD Flows Urban Network
Distribution Usage

Digital Traces (non-representative, high resolution)

Figure 8. Schematic view of the methods and data used to estimate an updated mode split for a city.

As a case study, we work with data from Santiago, Chile. The administrative units under study
are the several municipalities in the Santiago Metropolitan Area. We used the data sources de-
scribed in the previous section. Next, we describe the pre-processing and methods we apply to
these data.

Initial Guess of the Updated Mode Split

The first step in our methodology is to build an initial guess of an updated mode split for the city.
Recall that there are observations from official data that can aid an initial estimation of an updated
mode split, and we can use expert knowledge of the city as well. On the one hand, the National
Institute of Statistics (INE) provides a city-wide count of smart card transactions in the Metro system
and the number of car permits per municipality, as well as population projections. These statistics
are available for 2012 and 2020. On the other hand, we assume that taxi trips have increased, due
to the emergence of ride-hailing applications, although this increase is modest at commuting times
(Tirachini and Río, 2019); we also assume that pedestrian trips have decreased due to the effects
of the social outburst from October 2019. Taking these factors into account, and starting from the
trip counts per mode of transportation 𝑚 as measured by the Santiago Travel Survey 2012, we use
the updated population change from 2012 to 2020 and the aforementioned factors to estimate an
initial guess of the updated mode split in the following way:

√
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛20 (𝑚) 𝑀𝑒𝑡𝑟𝑜20
𝑀𝑇20 (𝑚) = 𝑀𝑇12 (𝑚) × × (1)
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛12 (𝑚) 𝑀𝑒𝑡𝑟𝑜12
√
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛20 (𝑚) 𝑃 𝑒𝑟𝑚𝑖𝑡𝑠20 (𝑚)
𝐶20 (𝑚) = 𝐶12 (𝑚) × × (2)
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛12 (𝑚) 𝑃 𝑒𝑟𝑚𝑖𝑡𝑠12 (𝑚)
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛20 (𝑚)
𝐴20 (𝑚) = 𝐴12 (𝑚) × × 0.975 (3)
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛12 (𝑚)
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛20 (𝑚)
𝑇20 (𝑚) = (𝑇12 (𝑚) + 1) × × 1.09. (4)
𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛12 (𝑚)
(5)

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 13 of 31
 In these equations, m is a municipality, Population is the corresponding municipal population
at the years 2012 and 2020, Metro is the number of smart card transactions, Permits is the cor-
responding municipal car permits at the years 2012 and 2020; MT is mass-transit, C is motorised
transport, A is active transport, and T is taxi. In all modes of transportation we multiply the number
of trips per municipality m with their ratio of change according to the population distribution. The
number of trips is the daily average for the period of study for 2020, to make it comparable to the
trip count obtained from the travel survey (i.e., an average commuting trip count for laboral days
in non-estival periods).
We adjusted MT according to available data: metro trips were reduced by 43.70%, however,
public transport includes buses and trains, and the metro network was severely affected by the
social outburst from 2019, with many stations closed or not operative during the period under
study. Hence, we modulate the change in smart card records using a square root. For motorised
trips, we adjusted C using a similar approach. In addition to updating the amount of trips according
to population change, we modulate them using the change in car permits per municipality. We also
modulate this rate of change with a square root, because not all cars are for commuting, as others
are for secondary usage at each household, and others are for work. Hence, we modulate the car
permit for those situations. Active transport A follows a similar schema. Here, we assume that
those trips have reduced their relative share by 2.5%, given the social context of the city. And
finally, for taxi trips T we applied the population change factor and the increase factor based on
previous reports on ride-hailing usage (Tirachini and Río, 2019). Note that we added one trip to the
observed taxi trip count in 2012 to avoid having zeroes in the updated modal partition, as some
municipalities did not have taxi trips in 2012 at commuting times in laboral days.
The ratio between this initial mode split estimation and a naive one where the original R01 is
multiplied municipality-wise with the fraction of population change is 0.9634. These equations are
fairly direct and new data sources or modulation factors can be introduced according to expert
knowledge if needed.

Data Integration through an Unified Representation

The datasets described earlier were generated by different sources, although they contain several
shared concepts. For instance, all of them have data related to the municipalities in the area under
study, either explicitly, or through aggregation of data. Here we describe a unified representation
that enables data integration.
First, we identified all different concepts present in the data (e.g., municipality, mode of trans-
portation, waypoint, etc.; see Table 1 for all of them, including description). Under this definition,
every table from the datasets expresses a relation between concepts as a matrix. For instance,
the mode split for different speed ranges (see Figure 3.a) is a matrix, where rows are modes of
transportation, columns are speed ranges, and cell values quantify the fraction of trips that corre-
spond to the specific mode/speed range pair. Analogously, the mode split per municipality is also
a matrix.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 14 of 31
 Table 1. List of concepts expressed in our datasets.

Concept Description Datasets Cardinality Example Values

Application Application or web domain DPI (2020) 963 [Link], [Link]

accessed by mobile phones.
Income Mean income at a given area, in EOD (2012), CASEN 5 Q01, Q05
quintiles. (2020)
Migration Country of origin of migrant CASEN (2020) 55 Spain, Venezuela
population.
Mode of Transportation Type of vehicle used for EOD (2012) 4 mass-transit, motorised
commuting trips.
Municipality Administrative unit used for All 40 Santiago, Providencia
analysis
Population Demographic characteristics INE (2012, 2020) 22 0 to 10 years, > 65 years
(age in decades) per year
Urban Infrastructure Characteristics of the built OSM (2020) 7 near railways, near
environment related to highways
transportation
Speed Average speed, aggregated in EOD (2012), XDR 8 < 5 Km/h, 30 to 60 Km/h
ranges (2020)
Waypoint A mobile phone tower DPI (2020), XDR 1878 BTS-0001, BTS-0087
(2020)
Work Type A profession or occupation CASEN (2020) 10 scientist, machine
operator

Second, we built a relationship network (see Figure 9) with all matrices under analysis (see Ta-
ble 2). We work with a total of 14 matrices referred to as Rn, where n ranges from 1 to 14. Most
of these matrices were built directly from the data. Note, however, that two of these matrices, Ap-
plication Usage and Mode of Transportation (R13), and Urban Infrastructure and Mode of Trans-
portation (R14) were manually built by us. For example, R13 was built relating apps domains such
as [Link] and [Link], with modes of transportation such as taxi and mass-transit . We
explored the full list of apps/domains to establish these associations and found 53 apps that could
be associated with at least one mode of transportation. Some domains are associated with two.
For instance, [Link] may be used by any motorised vehicle, and this includes the taxi category.
Additionally, the Municipality and Waypoints matrix (R05) was weighted using TF-IDF, a common
weighting method that normalises each row in the matrix. Previous experiments have shown that
such weighting increases the stability and accuracy of the model (Graells-Garrido et al., 2018b).
Supplementary Material Section A2 shows a visualisation of all relations in the dataset.
The association between domains and modes of transport goes from 0 (no assoc.) to 1 (full
assoc.). This leaves 1,007 apps without association. We assign a value of 0.25 to these (so all
associations for an app sum 1). Similarly, for the R14 matrix, the matrix which relate the urban
infrastructure of the city to a mode of transport, was manually assigned using a boolean flag (0
or 1) to each pair made of infrastructure type and mode of transportation, as described in the
OpenStreetMap section earlier.
The R01 matrix contains the initial solution of the mode split per municipality estimated earlier
in this section. In the remainder of this paper we describe a method to obtain an updated R01, as
the other matrices contain information that may help in improving the updated mode split.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 15 of 31
 Figure 9. Diagram of concept relations from our datasets. Nodes (circles) are the different concepts we
analyse from the data. Each edge is a relation between concepts, which we formally represent as a matrix
from the dataset. The purple arrow represents the mode split per municipality.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 16 of 31
 Table 2. All relations analysed. Each relation is a matrix built fromvthe data.

ID Source Concept Target Concept Cell Value Description Source

R01 Municipality Mode of Transportation The number of trips originated at a given Initial Estimation
municipality for a given mode of transportation. using EOD (2012)
and INE (2020)
R02 Municipality Work Types The number of people in a given profession that CASEN (2020)
lives in a given municipality.
R03 Municipality Migration The number of migrants from a given origin CASEN (2020)
country that live in a given municipality.
R04 Municipality Population The number of people in a given age range that INE (2012, 2020)
lives in a given municipality, for two different
years.
R05 Municipality Waypoints A weighted value estimated from the number of XDR (2020)
trips from people that live in a given
municipality that passes through a given
waypoint or mobile phone tower.
R06 Municipality Income The number of people in a given income CASEN (2020)
quintile that lives in a given municipality.
R07 Municipality Speed The number of trips originating at a given XDR (2020)
municipality that have an average speed within
a given speed range.
R08 Waypoints Speed The number of trips that pass through a given XDR (2020)
waypoint that have an average speed within a
given speed range.
R09 Waypoints Application Usage A weighted value estimated from the number of DPI (2020)
accesses to a given app/domain from a given
waypoint.
R10 Waypoints Urban Infrastructure Whether (boolean) there is a given type of XDR(2020) and
urban infrastructure in a ratio of 500 metres OSM (2020)
around a given waypoint.
R11 Income Mode of Transportation The fraction of trips in a given mode of EOD (2012)
transportation for a given income quintile.
R12 Speed Mode of Transportation The fraction of trips in a given mode of EOD (2012)
transportation for a given speed range.
R13 Application Usage Mode of Transportation The association (between 0 and 1) of a given Manually built
app/domain with a given mode of
transportation.
R14 Urban Infrastructure Mode of Transportation Whether (boolean) trips from a given mode of Manually built
transportation use a given urban infrastructure.

Data Fusion using Matrix Factorisation

At this point, we have a set of relationships between concepts. Although these relationships have a
common representation, as all of them are matrices, they have not been unified yet; in other words,
we do not get new insights from all these relationships being analysed together. In our context, we
would expect to improve our initial guess of mode split by employing the other relations expressed
in the data. Our goal is to compute a new R01 matrix (mode split) that contains an updated modal
split by using the information present in the rest of the dataset, which is up-to-date. We do so
through Data Fusion, a family of methods to integrate datasets into an unified representation with
the aim of extracting knowledge that cannot emerge from each data source alone (El Faouzi et al.,
2011).
In previous work, matrix factorisation has been used to infer mode of transportation usage
(Graells-Garrido et al., 2018b). One of the characteristics of matrix factorisation methods is the
ability to reconstruct matrices by means of matrix projection. A matrix 𝑅01′ could be expressed
as a reconstruction of 𝑅01 based on the other matrices available on the dataset. Previous work
using factorisation has employed Non-negative Matrix Factorisation (NMF), which decomposes a
positive matrix 𝑀𝑖𝑗 into the multiplication of two matrices:

𝐌𝐢𝐣 ≈ 𝐆𝐢 × 𝐆𝐣 , (6)

where 𝐺𝑖 is a low-rank latent representation of concept 𝑖 (rows of 𝑀𝑖𝑗 ), and 𝐺𝑗 is a low-rank latent
representation of concept 𝑗 (columns of 𝑀𝑖𝑗 ). In this schema, if the matrix to be decomposed is
𝑅01, then the latent dimensions are interpreted as modes of transportation (Graells-Garrido et al.,
2018b).

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 17 of 31
 Although NMF is widely used in multiple areas, it is insufficient for our problem, because there is
no direct way of incorporating the additional information available in the solution. In other words,
both concepts are clustered in the latent space without considering the network of relations. In
this situation, the Matrix Tri-factorization approach provides a definition of a co-clustering that
considers these connections (Ding et al., 2006; Žitnik and Zupan, 2014), by separating the latent
representation of each concept from the latent representation of their relationship:

𝐌𝐢𝐣 ≈ 𝐆𝐢 × 𝐒𝐢𝐣 × 𝐆𝐓𝐣 , (7)

where 𝑆𝑖𝑗 is denoted as the backbone matrix that encodes the relation 𝑀𝑖𝑗 in the latent space,
and 𝐺𝑖 (and, by analogy, 𝐺𝑗 ) is the latent representation of the concept 𝑖 in all its relationships. In
other words, the G matrices cluster concepts into latent dimensions, and the S matrices encode the
relationships between clusters of different concepts. The concept matrices have positive elements
only (as in NMF), whereas the backbone matrices may have negative elements. In this way, the
backbone matrix may express positive and negative associations between clusters of concepts.
Following this schema, the fully updated mode split matrix 𝑅01′ can be reconstructed from its
corresponding backbone matrix 𝑆, and the concept matrices 𝐺𝑚𝑜𝑑𝑒 and 𝐺𝑚𝑢𝑛𝑖𝑐𝑖𝑝𝑎𝑙𝑖𝑡𝑦 , which should be
computed taking into account the full network of relationships.
To perform the factorisation, the following problem must be solved:
∑ ‖ ‖
min ‖𝐌𝐢𝐣 − 𝐆𝐢 𝐒𝐢𝐣 𝐆𝐣 𝑇 ‖ . (8)
𝐆>=0 ‖ ‖
𝐌𝐢𝐣 ∈

Here, 𝑚𝑎𝑡ℎ𝑐𝑎𝑙𝑅 represents the set of relations in our data, 𝑚𝑎𝑡ℎ𝑏𝑓 𝑀𝑖𝑗 denotes a relation matrix
between concepts 𝑖 and 𝑗, and ‖⋅‖ represents the Frobenius norm. Although matrix factorisation is
an NP-HARD problem, an efficient implementation to solve this problem, based on multiplicative
updates (Žitnik and Zupan, 2014), is available in the scikit-fusion library (Čopar et al., 2019).
As in other factorisation methods, each concept matrix 𝐺𝑖 requires a rank parameter 𝑘𝑖 . The
rank 𝑘 determines the dimension of the latent space that represents the data; one interpretation
of its meaning is how much the data is compressed in its latent representation. For instance, if
𝑘 is very small in comparison to the total dimension of the concept, there is a risk of information
loss due to extreme compression; conversely, for large values of 𝑘, the model may overfit and give
more importance to noise than to the actual information in the model.
In our context, we aim at compressing the data enough to learn patterns behind each concept,
while at the same time allowing the reconstruction of the data to be different from its initial value,
as we assume that we are updating the data, rather than merely compressing/decompressing it.
Furthermore, we need to balance the complexity of the model (higher ranks imply a more complex
model) with the computational cost of performing the factorization. As such, we evaluated several
rank functions and followed the commonly used elbow method to select a function (see Supple-
mentary Material Section A3 for details). We selected the following heuristic to estimate the rank
𝑘𝑖 for a concept 𝑐𝑖 :
√
𝑘𝑖 = 2 × ∣ 𝑐𝑖 ∣ − 1, (9)

Where ∣ 𝑐𝑖 ∣ is the cardinality of 𝑐𝑖 (see Table 1).

By solving the optimization problem, all relations in the dataset can be reconstructed by multi-
plying the corresponding concept and backbone matrices. In particular, the reconstructed mode
split matrix 𝑅01′ can be obtained directly through the following multiplication:

𝐑𝟎𝟏′ = 𝐆𝐦𝐮𝐧𝐢𝐜𝐢𝐩𝐚𝐥𝐢𝐭𝐲 × 𝐒𝐦𝐮𝐧𝐢𝐜𝐢𝐩𝐚𝐥𝐢𝐭𝐲,𝐦𝐨𝐝𝐞 × 𝐆𝐓𝐦𝐨𝐝𝐞 . (10)

This 𝑅01′ matrix is the ultimate result of this pipeline, as it expresses the modal split in terms of
the knowledge available in the entire dataset, in addition to the initial estimation.
The solution to the optimization problem is guaranteed to be a local optimum only, due to
the method being based on multiplicative updates. Since this method employs random numbers

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 18 of 31
 to initialise and then improve a solution, the procedure to select the best model is to run multiple
instances, each with a different random seed for initialisation. The criteria for selection is based on
reconstruction error (Žitnik and Zupan, 2014). Here, we define the global error 𝑒 as the geometric
mean of normalised reconstruction error for all relations, defined as follows:
√
√ ‖ ‖
√ ‖𝐌𝐢𝐣 − 𝐆𝐢 𝐒𝐢𝐣 𝐆𝐣 𝑇 ‖
√ ‖ ‖.
𝑒= √
||
Π𝐌𝐢𝐣 ∈ (11)
‖ ‖
‖𝐌𝐢𝐣 ‖
‖ ‖
The model with the smallest reconstruction error 𝑒 will be selected. Note that the 𝑅01′ matrix is
excluded from this calculation as its difference with the original 𝑅01 matrix is not interpreted as an
error, rather, we interpret it as the change in modal split.
Next, we analyse the results of applying these methods to the data from Santiago, Chile.

Results
In this section we report the results of following the data fusion approach with data from Santiago,
Chile. First, we focus on the model selection procedure. Then, we analyse the updated mode split
reconstructed by the best model. And finally, we validate our results with an official data source.

Model Selection
In total, we adjusted 100 different models with different seeds for the random initialization. Here,
we report the reconstruction of the 𝑅01′ matrix in all these models (see Figure 10). We observe
that the relative mode share per municipality exhibits variability, however, the municipalities with
larger variability are outside of the main urban area. This could be expected, as peripheral munici-
palities have less data available. In general, the behaviour is arguably stable: the greatest standard
deviation in share change per mode of transportation is 0.12 (mass-transit), whereas the smallest
is 0.05 (taxi).
Regarding reconstruction error per matrix, there are several magnitudes depending on the
relation being analysed (see Figure 11). First, the matrices related to population statistics had a
very small error (median of 0.05 in all fitted models), whereas other matrices presented slightly
higher values (such as the median of 0.11 in the municipality and income relation). This implies
that the latent representations of the official data from the city are accurate. Moderate errors come
from matrices such as the relation between application usage and mode of transportation (median
of 0.44). This value was expected to be high, as few applications were known to be associated with
modes of transportation, and the remaining applications had a uniform imputed value of 0.25
(so they are equally associated with each mode of transportation). Larger errors come from the
mobile phone data relations (median of 0.87 for the municipality and waypoints relation), which are
larger and sparser than the other relations. As such, this was expected, as these relations exhibit
detailed behaviour from the population at waypoint resolution, whereas the other matrices, which
are used to reconstruct the data, contain data at a coarser resolution. We expect the model to find
the most important patterns connected to the other matrices, as in previous work using matrix
factorization of mobile phone data (Graells-Garrido et al., 2018b). To select the best model among
the 100 instances, we chose the one with the lowest global reconstruction error, defined as the
geometric mean of individual reconstruction errors. We observe that the municipality and mode
of transportation relation (𝑅01′ ) exhibited a median change of 0.2, which indicates a moderate
change in transportation patterns. Recall that this change is interpreted as the change in the modal
split rather than error. We discuss these changes for the selected model in the next section. Since
we focus the remainder of this paper in the updated mode split by this model and its validation,
we refer the reader to Supplementary Material Section A4 for a detailed view of all latent matrices
for concepts and relationships.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 19 of 31
 Figure 10. Each boxplot shows how the relative share of the mode split per municipality varies with respect
to 100 adjusted models. Each model has a different random initialization.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 20 of 31
 Figure 11. Boxplot of analysis of model error. Each glyph represents the normalised reconstruction error for
a relation in the dataset. A box comprises the range between the first and third quartile of values, as well as
the median of the distribution. Whisker lines express the spread of the data, and dots represent outlier
model instances. In total, we ran 100 model instances.

Updated Mode Split

Next, we used the best model to reconstruct an updated mode split for the city. Compared with
the initial mode split from 2012, we observe the following changes from the best factorisation
model: a decrease in mass-transit usage from 39.44% to 31.40%, an increase in motorised transport
from 40.64% to 50.36%, a decrease in active trips from 16.15% to 13.47%; and an increase in taxi
trips 3.76% to 4.76%. These changes are similar to those observed between the Santiago Travel
Survey 2012 and 2001: motorised transport increased by 5.1% and mass-transit was reduced by
6.4% (Muñoz et al., 2015).
The differences in the mode split reconstruction exhibit a geographical dependence on the dis-
tribution (see Figure 12). Most relevant changes happen in municipalities crossed or surrounded by
metro or train lines. Only three municipalities exhibit an increase in mass-transit share: Lo Espejo,
Lampa, and Peñaflor; of them, only Lo Espejo is in the urban area. The increase in mass-transit
share in Lo Espejo is expected since Pineda and Lira (2019) found that after the implementation of
the metro line 6 in 2017, trips with origin in Lo Espejo reduced their travel time by around 8 minutes.
Additionally, this municipality was benefited for the first time with a train station integrated into the
public transport system in 2017, whereas Lampa and Peñaflor may have been benefited by their
arguably close distance to the terminal stations of the new train and metro services put into oper-
ation in 2017 and 2019, respectively. Other municipalities exhibit a small variation in mass-transit,
including Cerrillos, San Joaquín, Estación Central, Lo Prado, Independencia, and Providencia. These
municipalities have access to new metro lines since 2017 (Pineda and Lira, 2019). Of particular in-
terest is the mass-transit increase in Vitacura, one of the wealthiest municipalities in the country,
with a strong share of motorised transport in the last travel survey. We hypothesise that, due to the
increase of remote work after the social outburst in 2019, the number of motorised trips decreased,
thus, increasing the relative share of mass-transit. We included a relationship between municipal-
ities and occupations to account for this phenomena in our dataset. However, as there is not an
official dataset regarding the share of remote work for 2022, the test of this hypothesis is left for
future work. Conversely, Conchalí, in the northern area of the urban radius, which is also crossed
by a new metro line, exhibited a high decrease in mass-transit share; in turn, all other three modes
increased considerably. Metro operations in Conchalí started in 2019; then, we hypothesise that
the social upheaval from that year delayed the expected modal shift to public [Link] other
municipalities decreased their mass-transit share. Most of them were already well-connected to
public transport. This difference could be explained by the poor quality of service of the public

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 21 of 31
 transport system, which is globally perceived as overcrowded and slow (Tirachini et al., 2017).

Figure 12. Geographical distribution of modal split differences per municipality.

Motorised, taxi and active trips exhibit an increase through municipalities in the entire area of
study. Arguably, the increase in motorised is the most expected relationship, as cars have increased
their sales in the last years (as evidenced by the increase in car permits), and as these vehicles are
likely to replace public transportation for their owners (Beirão and Cabral, 2007). A similar ex-
pectation holds with respect to taxi, although there are specific cases worthy of discussion. The
literature states that ride-hailing applications tend to replace public transport and traditional taxis
in Santiago (Tirachini and Río, 2019), and taxi exhibits a geographical pattern, as most of the mu-
nicipalities with increased taxi share belong to the Center, North and West macro areas of the city.
These macro areas had a stronger metro network than the others, thus, we may hypothesise that
ride-hailing has a tendency to replace metro trips in Santiago, including municipalities with access
to new metro lines. This result may be controversial, as it is expected that the number of trips in
public transportation would increase in those municipalities served by the new metro line since it
generated a decrease of 14% in travel time for its users (Pineda and Lira, 2019); thus, this opens
a line for future work regarding the effect of the social outburst in travel patterns in Santiago. Fi-
nally, one interesting pattern for us is how the municipalities where active transport increased are
primarily residential in the West and South macro areas. Further work is needed to understand
these changes, as this is an aspect of transportation affected by the COVID-19 pandemic and its
corresponding municipal mobility restrictions (Pappalardo et al., 2023).

Model Validation with External Sources

We made the first comparison between our results and ADATRAP at the macro area level (see
Table 3). We grouped the total averaged trips from each macro area. We observe that there is a
systematic difference, where the proposed model always predicted less trips than ADATRAP. This
could be related to our initial modulation of bip! smart card registers from the National Institute of

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 22 of 31
 Statistics, which could have been stronger, and thus, the model would have reported more mass-
transit trips. Then, to quantify how similar the prediction of our model was to the actual measured
trips at municipality level, we used the Pearson correlation coefficient 𝑟, defined as follows:
cov(𝑋, 𝑌 )
𝑟= . (12)
𝜎𝑋 𝜎𝑌

This coefficient measures the statistical association between two vectors with the municipal aver-
aged trip counts (𝑋 and 𝑌 ). The 𝑟 coefficient lies in [−1, 1], where a value of -1 indicates negative
correlation, a value of 1 indicates positive correlation, and 0 indicates total independence. The re-
sults obtained from the comparison yielded a value of 𝑟 = 0.88 (𝑝 < 0.001) for the proposed model,
suggesting a strong positive correlation between the predicted trips and the actual measured trips.
The initial guess of the updated mode split also had a correlation of 𝑟 = 0.84 (𝑝 < 0.001), indicating
that the initial estimation of mass-transit share was fairly accurate. The proposed model captures
behaviour observed in official data sources and improves the initial guess by integrating digital
traces and other official data.

Table 3. Comparison of average daily commuting trips between our proposed model and those reported by
the ADATRAP system.

Macro area Proposed model ADATRAP

Center 106,969 182,790

North 80,261 107,999
East 129,314 184,268
West 137,696 212,244
South 190,424 244,946
Total 644,664 932,247

Figure 13. The bubble plot of correlation shows the relationship between the mass-transit trip counts
predicted by the proposed model and the actual counts observed in smart card data for the same period of
study. Each bubble in the plot represents a municipality and its size is proportional to the municipality’s
population in 2020. In addition, each bubble is connected to a small dot that represents the initial mass-transit
trip count used by the model.

On the other hand, since we integrated several official data sources, where some of them are
not directly related to transportation, we tested two additional data configurations: one without
DPI data, and without any mobile phone data (XDR and DPI), with all the other sources remaining

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 23 of 31
 active in each configuration. Note that the process to adjust a model in each configuration is always
the same, the only difference is the amount of available relation matrices. For each configuration,
we fitted 100 model instances, and then we estimated the Pearson correlation coefficient 𝑟 between
each instance and the observed data by ADATRAP (see Figure 14). We observed that the global
error is always greater when using more data. This can be explained due to the size and density of
the relationships that involve waypoints (mobile phone towers), which are one order of magnitude
greater than the rest of the data. We also observe that our criteria to select the best model does
not always pick the model with the best correlation with ADATRAP data. In the configuration with
all data available, the best model is one of the most correlated with observed data. However, in
the other configurations, this is not guaranteed; indeed, in the configuration without mobile phone
data, the best model has one of the worst correlations with observed data. Additionally, the best
models for the test configurations present a worse result than the baseline correlation obtained
with the initial guess of the mode split. Thus, even though our best model criteria does not find the
actual best model according to ground truth data, it does pick one that is considerably better than
the baseline. This suggests that future work should identify an ad-hoc strategy for model selection.

Figure 14. Relationship between model error and correlation between model outcomes for mass transit and
validation data. We tested three data configurations: the proposed one that fuses all available data (official
data, OpenStreetMap and mobile phone data from XDR and DPI), one with the same data with exception of
DPI (application usage), and one using only official data and OSM. Each marker represents a model within
each group, with the best model of each configuration highlighted with a greater size and with colour.

Finally, one of the contributions from this work is the usage of DPI (application usage) from
mobile phones. Then, it would be important to assess the effect of incorporating this source into
the model. The existence of a potentially better or comparable result using less data requires addi-
tional analysis to justify the usage of several types of mobile phone data. At this point, we already
know that one quality is its determinism, that is, being able to select the best model using an estab-
lished criteria based on global error. However, we also need to identify how different the results
are. To do so, we estimated the Pearson correlation coefficients for each mode split between our
proposal and the other two data configurations (see Table 4). We observe that the mode splits
are similar between configurations with mobile phone data, with high correlation coefficients (all

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 24 of 31
 greater or equal than 0.9, with statistical significance). We also observe that the only mode of trans-
portation that presents a divergence from the obtained mode splits is taxi when working without
mobile phone data (0.49, 𝑝 = 0.001). This has a two-fold implication. On the one hand, it becomes
clear that the contribution of the mobile phone data data is the ability to pinpoint the new taxi
share, arguably through the signal of ride-hailing application usage. For all other modes, the of-
ficial data used may have latent factors that enable the model to identify part of their changes,
however, such analysis should be approached in future work. On the other hand, although this
result suggests that without using DPI (app usage), or even XDR (trajectories), we found that it is
possible to have a good up-to-date estimation of the modal partition. We highlight that the two
types of mobile phone data are necessary, as only when using them we have a replicable criteria
to find a model with good results in comparison to the baseline comparison.

Table 4. Pearson correlation coefficients between the predicted mode split in the proposed data
configuration and two configurations with less data but similar validation results. All p-values are
Bonferroni-corrected.

No DPI No Mobile Phone Data

Mass Transit 0.98 (p < 0.001) 0 .97 (p < 0.001)

Motorised 0.98 (p < 0.001) 0 .97 (p < 0.001)
Taxi 0.90 (p = 0.4) 0.49 (p = 0.001)
Active 0.95 (p < 0.001) 0 .93 (p < 0.001)

Discussion
In practice, the Data Fusion approach requires a limited amount of data (a weekday of mobile
phone data, as in the case study) and supports incorporating domain knowledge or any other
data source that can be expressed as a matrix. This is a strength of the proposed method, as the
demand for frameworks to fuse data from a multidisciplinary perspective is increasing (Meta et al.,
2022), and incorporating domain knowledge is crucial to the adoption of data-driven initiatives for
transportation planning (Graells-Garrido et al., 2020). Another strength is that the model works
with aggregated data. Although we used individual-level mobile phone data, the actual input for
the model was mobility behaviour aggregated at administrative levels (for XDR trajectory data), and
at cell phone tower levels (for DPI application usage data). The usage of aggregated data arguably
diminishes potential issues regarding privacy, and facilitates data transfer between entities, due
to its reduced size and its inherent anonymisation. Furthermore, the good correlations observed
with models using different subsets of the available data indicate that the data fusion approach
has potential even when using official data and OSM only, which also opens a future line of work to
identify which already available official sources can aid in the determination of an updated mode
split.
The main theoretical implication of this work is that data-driven transportation planning projects
can still provide relevant insights even in the presence of biassed data. Data bias is commonly con-
sidered an artefact affecting digital traces, such as the mobile phone data integrated here. Such
bias comes from access to technology due to income, cultural, social, or geographical barriers.
Note that the model implicitly accounts for this because the integration of official socio-economic
data affects how the model computes the representation of each concept in the dataset: the mo-
bile phone data is used as one additional data source, not the main one in estimating the mode
split. We find this relevant because assessing bias for mobile phone data is challenging, due to the
data being anonymised. Indeed, despite the presence of bias in the digital traces, the correlation
obtained by our model is high and statistically significant.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 25 of 31
 Limitations
We identify three main limitations in this work. First, our validation only concerns the public trans-
port share, although the observed differences in other modes of transportation are coherent with
expectations from the literature and the urban context. However, to assess all the model out-
comes, a comparison with official data that includes all the other modes of transportation must be
made. To the extent of our knowledge, the data necessary to make such a comparison does not
exist at the moment of writing this paper.
Second, our method integrates official data regarding socio-economic characteristics, occupa-
tions, migration, and the local population. Arguably these sources implicitly control for data bias,
yet the extent of such control is not quantified. This aspect is relevant mainly for policymakers and
urban planners, as they would base their work on the results reported by data-driven models such
as the one presented here. We also observe that the temporal granularity of the data we analysed
is different, however, we argue that the phenomena under study (mode split) is depicted with the
finest detail possible, as we are using mobile phone data, whereas official sources may represent
yearly averages of phenomena that is well captured by the official sources as, indeed, the authority
and other institutions make use of that data every year.
Third, our results are bound by the mobile phone network event frequency and the geograph-
ical tower distribution. Connected phones could be redirected between towers if this is needed to
handle network requests, which could result in false positives in trip detection. However, XDR is
an aggregated dataset that consolidates multiple connections into a single record, with an interval
of 15 minutes in particular in our context. Although we cannot control by this problem, as the XDR
data is already aggregated, we expect its effect to be negligible in our setting due to the consoli-
dation of the data, the fact that in our study period the expected behaviour of people is to travel
to work or study locations, and that towers are distributed in the city to ensure quality of service
(as reported by the mobile phone operator). As a result of this limitation, a direct application of
this work into other countries or datasets may need a sensitivity analysis to determine if mobility
patterns inferred from XDR (or similar) have enough resolution to be included in the Data Fusion
model.

Future Work
There is still a space to be explored regarding the variability of outcomes given different initial con-
ditions, as well as the solution space of the proposed method. On the one hand, the initial solution
used as base for the Data Fusion process provided a solid starting point, however, the method is
bound by these initial conditions; a different strategy to build the initial solution may provide other
results. Yet, our validation proves that this was, indeed, a well defined initial condition. On the other
hand, our process to select a model did not perform any regularisation or reduction of the solution
space, as it was based only on the random initialization of the latent matrices. One potential way to
subvert this limitation is to include an additional operation in the optimization problem: restriction
of must-link and cannot-link values of each concept. This operation is supported by the original
model (Žitnik and Zupan, 2014). However, the specification of these restrictions needs additional
domain knowledge. We leave the exploration of that solution to future work.
Our work focuses on commuting times, the most recurring activity in a city, however, there are
many other activities and time periods that should be accounted for in urban planning. We expect
that the model could be made temporally-aware by working with tensors in all relationships that
may have a temporal dimension, such as the aggregated trajectories or the application usage per
municipality (Chen et al., 2019).

Conclusions
In this study, we developed a novel methodology to address the data challenges in evaluating and
informing transport planning interventions in cities. By utilising mobile phone data and data fusion

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 26 of 31
 techniques, we accurately estimated the mode of transportation usage in Santiago, a major Latin
American city.
Our findings reveal significant shifts in transportation modes over time. We observed a de-
crease in mass-transit usage, except in areas with new metro/rail lines, indicating the impact of
these infrastructure improvements. Motorised transport has increased citywide, highlighting the
ongoing challenges in promoting sustainability. Additionally, taxi usage, including traditional taxis
and ride-hailing applications, has risen, while the share of active transportation, such as walking
and cycling, has declined. We validated these results using official data from smart card trans-
actions, demonstrating the reliability and accuracy of our methodology. The consistency of our
findings with domain knowledge and historical trends further supports the robustness of our ap-
proach.
Looking ahead, as cities face escalating challenges in achieving sustainable transportation goals,
cost-effective tools providing finer data resolutions will be essential. Collaboration among scien-
tists, private and public institutions, and the continued development of innovative methods, like
ours, will play a pivotal role in addressing these challenges effectively.
In conclusion, our research advances the understanding of travel behaviour patterns and un-
derscores the need for data-driven approaches in sustainable transport planning. By harnessing
mobile phone data and data fusion techniques, cities can make informed decisions to promote
public transit, reduce reliance on motorised vehicles, and foster a more sustainable and resilient
urban future.

References
Acar, C., Dincer, I., 2020. The potential role of hydrogen as a sustainable transportation fuel to
combat global warming. International Journal of Hydrogen Energy 45, 3396–3406.
Allen, S.K., Barros, V., Burton, I., Campbell-Lendrum, D., Cardona, O.-D., Cutter, S.L., Dube, O.P., Ebi,
K.L., Field, C.B., Handmer, J.W., al., et, 2012. Summary for policymakers, in: Field, C.B., Barros,
V., Stocker, T.F., Dahe, Q. (Eds.), Managing the Risks of Extreme Events and Disasters to Advance
Climate Change Adaptation: Special Report of the Intergovernmental Panel on Climate Change.
Cambridge University Press, pp. 3–22. [Link]
Antoniou, C., Balakrishna, R., Koutsopoulos, H.N., 2011. A synthesis of emerging data collection
technologies and their impact on traffic management applications. European Transport Re-
search Review 3, 139–148.
Asgari, F., 2016. Inferring user multimodal trajectories from cellular network metadata in metropoli-
tan areas (PhD thesis). Institut National des Télécommunications.
Bachir, D., Khodabandelou, G., Gauthier, V., El Yacoubi, M., Puchinger, J., 2019. Inferring dynamic
origin-destination flows by transport mode using mobile phone data. Transportation Research
Part C: Emerging Technologies 101, 254–275. [Link]
Bantis, T., Haworth, J., 2017. Who you are is how you travel: A framework for transportation mode
detection using individual and environmental characteristics. Transportation Research Part C:
Emerging Technologies 80, 286–309.
Beirão, G., Cabral, J.S., 2007. Understanding attitudes towards public transport and private car: A
qualitative study. Transport policy 14, 478–489.
Breyer, N., Rydergren, C., Gundlegård, D., 2022. Semi-supervised mode classification of inter-city
trips from cellular network data. Journal of Big Data Analytics in Transportation 1–17.
Browning, R.C., Baker, E.A., Herron, J.A., Kram, R., 2006. Effects of obesity and sex on the energetic
cost and preferred speed of walking. Journal of applied physiology 100, 390–398.
Brueckner, J.K., 2005. Transport subsidies, system choice, and urban sprawl. Regional Science and
Urban Economics 35, 715–733.
Calabrese, F., Colonna, M., Lovisolo, P., Parata, D., Ratti, C., 2011. Real-time urban monitoring using
cell phones: A case study in rome. IEEE Transactions on Intelligent Transportation Systems 12,

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 27 of 31
 141–151. [Link]
Chang, X., Wu, J., Liu, H., Yan, X., Sun, H., Qu, Y., 2019. Travel mode choice: A data fusion model
using machine learning methods and evidence from travel diary survey data. Transportmetrica
A: Transport Science 15, 1587–1612. [Link]
Chen, K.-Y., Shah, R.C., Huang, J., Nachman, L., 2017. Mago: Mode of transport inference using the
hall-effect magnetic sensor and accelerometer. Proceedings of the ACM on Interactive, Mobile,
Wearable and Ubiquitous Technologies 1, 1–23.
Chen, X., He, Z., Chen, Y., Lu, Y., Wang, J., 2019. Missing traffic data imputation and pattern dis-
covery with a bayesian augmented tensor factorization model. Transportation Research Part C:
Emerging Technologies 104, 66–77.
Colvile, R., Hutchinson, E.J., Mindell, J., Warren, R., 2001. The transport sector as a source of air
pollution. Atmospheric environment 35, 1537–1565.
Čopar, A., Zupan, B., Zitnik, M., 2019. Fast optimization of non-negative matrix tri-factorization.
PloS one 14, e0217994.
Dabiri, S., Heaslip, K., 2018. Inferring transportation modes from GPS trajectories using a convolu-
tional neural network. Transportation research part C: emerging technologies 86, 360–371.
Ding, C., Li, T., Peng, W., Park, H., 2006. Orthogonal nonnegative matrix t-factorizations for cluster-
ing, in: Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, KDD ’06. Association for Computing Machinery, New York, NY, USA, pp. 126–
135. [Link]
Doyle, J., Hung, P., Kelly, D., McLoone, S.F., Farrell, R., 2011. Utilising mobile phone billing records
for travel mode discovery.
El Faouzi, N.-E., Leung, H., Kurian, A., 2011. Data fusion in intelligent transportation systems:
Progress and challenges–a survey. Information Fusion 12, 4–10.
García, P., Herranz, R., Javier, J., 2016. Big data analytics for a passenger-centric air traffic manage-
ment system. 6th SESAR Innovation Days.
González, M.C., 2019. Mining urban lifestyles: Urban computing, human behavior and recom-
mender systems. Computing. [Link]
Graells-Garrido, E., Caro, D., Miranda, O., Schifanella, R., Peredo, O.F., 2018a. The WWW (and an
H) of mobile application usage in the city: The what, where, when, and how, in: Companion
Proceedings of the the Web Conference 2018. pp. 1221–1229.
Graells-Garrido, E., Caro, D., Parra, D., 2018b. Inferring modes of transportation using mobile
phone data. EPJ Data Science 7, 49.
Graells-Garrido, E., Ferres, L., Caro, D., Bravo, L., 2017. The effect of Pokémon Go on the pulse of
the city: A natural experiment. EPJ Data Science 6, 1–19.
Graells-Garrido, E., Peña-Araya, V., Bravo, L., 2020. Adoption-driven data science for transportation
planning: Methodology, case study, and lessons learned. Sustainability 12. [Link]
su12156001
Graells-Garrido, E., Peredo, O., García, J., 2016. Sensing urban patterns with antenna mappings:
The case of Santiago, Chile. Sensors 16, 1098.
Graells-Garrido, E., Schifanella, R., Opitz, D., Rowe, F., 2022. Measuring the local complementarity
of population, amenities and digital activities to identify and understand urban areas of interest.
Environment and Planning B: Urban Analytics and City Science 23998083221117830.
Grantz, K.H., Meredith, H.R., Cummings, D.A., Metcalf, C.J.E., Grenfell, B.T., Giles, J.R., Mehta, S.,
Solomon, S., Labrique, A., Kishore, N., others, 2020. The use of mobile phone data to inform
analysis of COVID-19 pandemic epidemiology. Nature communications 11, 4961.
Green, M., Pollock, F.D., Rowe, F., 2021. New forms of data and new forms of opportunities to mon-
itor and tackle a pandemic, in: COVID-19 and Similar Futures: Pandemic Geographies. Springer,
pp. 423–429.
Gschwender, A., Munizaga, M., Simonetti, C., 2016. Using smart card and GPS data for policy and
planning: The case of Transantiago. Competition and Ownership in Land Passenger Transport

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 28 of 31
 (selected papers from the Thredbo 14 conference) 59 IS -, 242–249.
Haklay, M., 2010. How good is volunteered geographical information? A comparative study of
OpenStreetMap and ordnance survey datasets. Environment and planning B: Planning and
design 37, 682–703.
Heinen, E., Chatterjee, K., 2015. The same mode again? An exploration of mode choice variability
in great britain using the national travel survey. Transportation Research Part A: Policy and
Practice 78, 266–282.
Holleczek, T., The Anh, D., Yin, S., Jin, Y., Antonatos, S., Goh, H.L., Low, S., Shi-Nash, A., 2015. Traffic
measurement and route recommendation system for mass rapid transit (mrt), in: Proceedings
of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.
pp. 1859–1868.
Horn, C., Gursch, H., Kern, R., Cik, M., 2017. QZTool—automatically generated origin-destination
matrices from cell phone trajectories, in: Advances in Human Aspects of Transportation: Pro-
ceedings of the AHFE 2016 International Conference on Human Factors in Transportation, July
27-31, 2016, Walt Disney World®, Florida, USA. Springer, pp. 823–833.
Horn, C., Kern, R., 2015. Deriving public transportation timetables with large-scale cell phone data.
Procedia Computer Science 52, 67–74.
Huang, H., Cheng, Y., Weibel, R., 2019. Transport mode detection based on mobile phone network
data: A systematic review. Transportation Research Part C: Emerging Technologies 101, 297–
312.
Huang, Z., Ling, X., Wang, P., Zhang, F., Mao, Y., Lin, T., Wang, F.-Y., 2018. Modeling real-time human
mobility based on mobile phone and transportation data fusion. Transportation research part
C: emerging technologies 96, 251–269.
Hui, K.T.Y., Wang, C., Kim, A., Qiu, T.Z., 2017. Investigating the use of anonymous cellular phone
data to determine intercity travel volumes and modes.
Instituto Nacional de Estadísticas, 2022. Estadísticas.
Jain, J., Lyons, G., 2008. The gift of travel time. Journal of transport geography 16, 81–89.
Kalatian, A., Shafahi, Y., 2016. Travel mode detection exploiting cellular network data, in: MATEC
Web of Conferences. EDP Sciences, p. 03008.
Karlaftis, M.G., 2004. A DEA approach for evaluating the efficiency and effectiveness of urban tran-
sit systems. European Journal of Operational Research 152, 354–364.
Kuhnimhof, T., Chlond, B., Von Der Ruhren, S., 2006. Users of transport modes and multimodal
travel behavior: Steps toward understanding travelers’ options and choices. Transportation
research record 1985, 40–48.
Lee, Y., Circella, G., Mokhtarian, P.L., Guhathakurta, S., 2020. Are millennials more multimodal? A
latent-class cluster analysis with attitudes and preferences among millennial and generation x
commuters in california. Transportation 47, 2505–2528.
Li, G., Chen, C.-J., Peng, W.-C., Yi, C.-W., 2017. Estimating crowd flow and crowd density from cel-
lular data for mass rapid transit, in: Proceedings of the 6th International Workshop on Urban
Computing, Halifax, NS, Canada. pp. 18–30.
Li, Y., Wu, F.-X., Ngom, A., 2018. A review on machine learning principles for multi-view biological
data integration. Briefings in bioinformatics 19, 325–340.
Meta, I., Cucchietti, F.M., Navarro-Mateu, D., Graells-Garrido, E., Guallart, V., 2022. A physiology-
inspired framework for holistic city simulations. Cities 126, 103553. [Link]
2021.103553
Ministerio de Desarrollo Social y Familia, 2020. Encuesta CASEN en Pandemia 2020.
Molin, E., Mokhtarian, P., Kroesen, M., 2016. Multimodal travel groups and attitudes: A latent class
cluster analysis of dutch travelers. Transportation Research Part A: Policy and Practice 83, 14–
29.
Monroe, B.L., Colaresi, M.P., Quinn, K.M., 2008. Fightin’words: Lexical feature selection and evalu-
ation for identifying the content of political conflict. Political Analysis 16, 372–403.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 29 of 31
 Munizaga, M.A., Gschwender, A., Gallegos, N., 2020. Fare evasion correction for smartcard-based
origin-destination matrices. Transportation Research Part A: Policy and Practice 141, 307–322.
Munizaga, M.A., Palma, C., 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.
Muñoz, V., Thomas, A., Navarrete, C., Contreras, R., 2015. Encuesta origen-destino de Santiago
2012: Resultados y validaciones. Estudios de Transporte 19.
Pappalardo, L., Cornacchia, G., Navarro, V., Bravo, L., Ferres, L., 2023. A dataset to assess mobility
changes in chile following local quarantines. Scientific Data 10, 6.
Phithakkitnukoon, S., Sukhvibul, T., Demissie, M., Smoreda, Z., Natwichai, J., Bento, C., 2017. In-
ferring social influence in transport mode choice using mobile phone data. EPJ Data Science 6,
1–29.
Pineda, C., Lira, B.M., 2019. Travel time savings perception and well-being through public transport
projects: The case of Metro de Santiago. Urban Science 3, 35.
Qu, Y., Gong, H., Wang, P., 2015. Transportation mode split with mobile phone data, in: Intelligent
Transportation Systems (ITSC), 2015 IEEE 18th International Conference On. IEEE, pp. 285–289.
Rodríguez-Núñez, E., García-Palomares, J.C., 2014. Measuring the vulnerability of public transport
networks. Journal of transport geography 35, 50–63.
Rowe, F., 2023. Big data, in: Concise Encyclopedia of Human Geography. Edward Elgar Publishing,
pp. 42–47.
Rowe, F., 2022. Using digital footprint data to monitor human mobility and support rapid humani-
tarian responses. Regional Studies, Regional Science 9, 665–668.
Rowe, F., Arribas-Bel, D., Calafiore, A., MacDonald, J., Samardzhiev, K., Fleischmann, M., 2022. Mo-
bility data in urban science. The Alan Turing Institute.
Rowe, F., Bell, M., 2020. The drivers of long-distance commuting in Chile: The role of the spatial
distribution of economic activities, in: Population Change and Impacts in Asia and the Pacific.
Springer, pp. 123–149.
Schlaich, J., Otterstätter, T., Friedrich, M., others, 2010. Generating trajectories from mobile phone
data, in: Proceedings of the 89th Annual Meeting Compendium of Papers, Transportation Re-
search Board of the National Academies.
SECTRA, 2015. Informe ejecutivo encuesta origen-destino de viajes Santiago.
Semanjski, I., Gautama, S., Ahas, R., Witlox, F., 2017. Spatial context mining approach for transport
mode recognition from mobile sensed big data. Computers, Environment and Urban Systems
66, 38–52.
Smoreda, Z., Olteanu-Raimond, A.-M., Couronné, T., 2013. Spatiotemporal data from mobile phones
for personal mobility assessment, in: Transport Survey Methods: Best Practice for Decision
Making. Emerald Group Publishing Limited.
Somma, N.M., Bargsted, M., Disi Pavlic, R., Medel, R.M., 2021. No water in the oasis: The Chilean
Spring of 2019–2020. Social Movement Studies 20, 495–502.
Suazo-Vecino, G., Muñoz, J.C., Fuentes Arce, L., 2019. The displacement of Santiago de Chile’s down-
town during 1990–2015: Travel time effects on eradicated population. Sustainability 12, 289.
Tenkanen, H., contributors, pyrosm, 2022. pyrosm: OpenStreetMap PBF data parser for Python.
Thomson, J.M., 1998. Reflections on the economics of traffic congestion. Journal of transport eco-
nomics and policy 93–112.
Tirachini, A., Hurtubia, R., Dekker, T., Daziano, R.A., 2017. Estimation of crowding discomfort in
public transport: Results from Santiago de Chile. Transportation Research Part A: Policy and
Practice 103, 311–326.
Tirachini, A., Río, M. del, 2019. Ride-hailing in Santiago de Chile: Users’ characterisation and effects
on travel behaviour. Transport Policy 82, 46–57.
Urquieta Ch., C., 2019. Dura pérdida para el metro: No tiene seguros para estaciones ni trenes.
CIPER.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv | 30 of 31
 Vuchic, V.R., 2007. Urban transit systems and technology. John Wiley & Sons.
Wang, H., Calabrese, F., Di Lorenzo, G., Ratti, C., 2010. Transportation mode inference from anonymized
and aggregated mobile phone call detail records, in: Intelligent Transportation Systems (ITSC),
2010 13th International IEEE Conference On. IEEE, pp. 318–323.
Wu, W., Cheu, E.Y., Feng, Y., Le, D.N., Yap, G.E., Li, X., 2013. Studying intercity travels and traffic
using cellular network data. Mobile Phone Data for Development: Net Mob 2013.
Zannat, K.E., Choudhury, C.F., 2019. Emerging big data sources for public transport planning: A
systematic review on current state of art and future research directions. Journal of the Indian
Institute of Science 99, 601–619.
Zhang, L., Pfoser, D., 2019. Using OpenStreetMap point-of-interest data to model urban change—a
feasibility study. PloS one 14, e0212606.
Žitnik, M., Zupan, B., 2014. Data fusion by matrix factorization. IEEE transactions on pattern analy-
sis and machine intelligence 37, 41–53.

Graells-Garrido et al. 2023 | Updating Mode of Transportation Distributions from Travel Surveys using Data Fusion arXiv