Geo-located Twitter as proxy for global mobility patterns
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
Hawelka, Bartosz, Izabela Sitko, Euro Beinat, Stanislav
Sobolevsky, Pavlos Kazakopoulos, and Carlo Ratti. “GeoLocated Twitter as Proxy for Global Mobility Patterns.”
Cartography and Geographic Information Science 41, no. 3
(February 26, 2014): 260–271.
As Published
http://dx.doi.org/10.1080/15230406.2014.890072
Publisher
Cartography & Geographic Information Society
Version
Final published version
Accessed
Thu May 26 18:50:24 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/101644
Terms of Use
Creative Commons Attribution
Detailed Terms
http://creativecommons.org/licenses/by/4.0/
Cartography and Geographic Information Science
ISSN: 1523-0406 (Print) 1545-0465 (Online) Journal homepage: http://www.tandfonline.com/loi/tcag20
Geo-located Twitter as proxy for global mobility
patterns
Bartosz Hawelka, Izabela Sitko, Euro Beinat, Stanislav Sobolevsky, Pavlos
Kazakopoulos & Carlo Ratti
To cite this article: Bartosz Hawelka, Izabela Sitko, Euro Beinat, Stanislav Sobolevsky,
Pavlos Kazakopoulos & Carlo Ratti (2014) Geo-located Twitter as proxy for global
mobility patterns, Cartography and Geographic Information Science, 41:3, 260-271, DOI:
10.1080/15230406.2014.890072
To link to this article: http://dx.doi.org/10.1080/15230406.2014.890072
© 2014 The Author(s). Published by Taylor &
Francis.
Published online: 26 Feb 2014.
Submit your article to this journal
Article views: 3286
View related articles
View Crossmark data
Citing articles: 35 View citing articles
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=tcag20
Download by: [18.51.1.88]
Date: 09 March 2016, At: 07:52
Cartography and Geographic Information Science, 2014
Vol. 41, No. 3, 260–271, http://dx.doi.org/10.1080/15230406.2014.890072
Geo-located Twitter as proxy for global mobility patterns
a,b
Bartosz Hawelka *, Izabela Sitkoa,b, Euro Beinata, Stanislav Sobolevskyb, Pavlos Kazakopoulosa and Carlo Rattib
a
Department of Geoinformatics – Z_GIS, GISscience Doctoral College, University of Salzburg, Salzburg, Austria; bSENSEable
City Laboratory, Massachusetts Institute of Technology, Cambridge, MA, USA
(Received 17 October 2013; accepted 30 December 2013)
Downloaded by [] at 07:52 09 March 2016
Pervasive presence of location-sharing services made it possible for researchers to gain an unprecedented access to the
direct records of human activity in space and time. This article analyses geo-located Twitter messages in order to uncover
global patterns of human mobility. Based on a dataset of almost a billion tweets recorded in 2012, we estimate the volume
of international travelers by country of residence. Mobility profiles of different nations were examined based on such
characteristics as mobility rate, radius of gyration, diversity of destinations, and inflow–outflow balance. Temporal patterns
disclose the universally valid seasons of increased international mobility and the particular character of international travels
of different nations. Our analysis of the community structure of the Twitter mobility network reveals spatially cohesive
regions that follow the regional division of the world. We validate our result using global tourism statistics and mobility
models provided by other authors and argue that Twitter is exceptionally useful for understanding and quantifying global
mobility patterns.
Keywords: geo-located Twitter; global mobility patterns; community detection; collective sensing
Introduction
Reliable and effective monitoring of worldwide mobility
patterns plays an important role in studies exploring migration flows (Castles and Miller 1998; Greenwood 1985;
Sassen 1999) and tourist activity (Miguéns and Mendes
2008), as well as for examining the spread of diseases
and for epidemic modeling (Bajardi et al. 2011; Balcan
et al. 2009). Traditionally, those studies relied either on
aggregated and temporally sparse official statistics or on
selective, small-scale observations and surveys. In a more
recent approach, worldwide mobility was approximated
using air traffic volumes (Barrat et al. 2004). However,
this dataset of potentially global coverage is biased toward
just one mode of transportation and is, in many cases,
difficult to obtain. Data about location-sharing services
provide researchers with much more accurate records of
human activity. Each day, millions of individuals leave
behind digital traces of their activities by using mobile
phones, credit cards, or social media. Most of those traces
can be located in space and time, and thus they constitute a
valuable source for human mobility studies.
Out of several types of collectively sensed data, cellular
phone records have been the most intensively explored for
analysis of human mobility. Because cell phones are almost
universally used, data about their use have led to important
findings about movement in urban (Calabrese et al. 2010;
Kang et al. 2012), regional (Calabrese et al. 2013; Sagl
et al. 2012), and national scales (Krings et al. 2009; Simini
et al. 2012). High fragmentation of the mobile telecom
market, however, precludes the availability of a worldwide
dataset on cell phone use. In this case, social media data are
a good alternative. Despite its lower penetration and a
potential bias toward younger populations, social media’s
popularity and representativeness are high and growing
(Gesenhues 2013), and in most cases the media are global
by design.
In this study, we attempt to uncover global mobility
patterns and compare mobility characteristics of different
nations. Our work is based on the data from Twitter – one
of the most popular social media platforms, with over 500
million users registered by mid-2013 (Twitter Statistics
2013). Initially established in the USA, the service has
quickly spread to other countries (Java et al. 2007; Leetaru
et al. 2013; Mocanu et al. 2013), becoming a worldwide
phenomenon. By design, Twitter is an open and public
medium, which practically limits privacy consideration,
especially in studies such as ours which examine collective rather than individual patterns of human behavior. In
particular, we take advantage of tweets augmented with
explicit geographic coordinates provided by either GPS
embedded in a mobile device or the IP address of a
computer. These geo-located tweets account for around
1% of the total feed (Morstatter et al. 2013). However,
thanks to the increasing penetration of smart devices and
mobile applications, the volume of the geo-located Twitter
has been constantly growing (Figure 1), becoming an
*Corresponding author. Email: bartosz.hawelka@sbg.ac.at
© 2014 The Author(s). Published by Taylor & Francis.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Cartography and Geographic Information Science
4M
75M
3M
50M
2M
25M
1M
December
October
Users
November
September
July
August
June
May
April
March
February
January
Tweets
# of users
# of tweets
100M
Downloaded by [] at 07:52 09 March 2016
Figure 1. Number of geo-located tweets (blue line) and users
(orange line) per month in 2012.
invaluable register of human traces in space and time. The
absolute volume of 3.5M geo-located tweets per day
(authors’ calculation for December 2012) appears as a
promising base for carrying out a worldwide mobility
analysis, which is the objective of our exploration.
Because of richness yet simplicity of the medium,
Twitter has already been the subject of many studies for
a variety of applications. First explorations focused on the
properties of Twitter as a social network, proving its
global character and scientific potential as early as one
year after its launch (Huberman, Romero, and Wu 2008;
Java et al. 2007; Kwak et al. 2010). Another line of
research examined the content of tweets to assess society’s
mood (Bollen, Pepe, and Mao 2011; Golder and Macy
2011; Pak and Paroubek 2010). Recently, this research
was done also with a geographic perspective (Frank
et al. 2013; Mitchell et al. 2013). Yet another area that
has received much attention is crisis management
(MacEachren et al. 2011; Sakaki, Okazaki, and Matsuo
2010; Thom et al. 2012), where the emphasis was on the
detection of anomalous activity and the potential of a
locally generated content to inform emergency services.
Geo-located Twitter data have been used to inform urban
management and planning (Frias-Martinez et al. 2012;
Wakamiya, Lee, and Sumiya 2011), as well as in public
health assessment (Ghosh and Guha 2013). All the aforementioned studies were spatially selective, focusing on
specific study areas. The global perspective was introduced in the study of Kamath et al. (2012), with the
analysis of the geographic spread of hashtags. Leetaru
et al. (2013) attempted to describe the geography of
Twitter based on a one-month sample of global geolocated tweets, while Mocanu et al. (2013) described the
global distribution of the different languages used while
tweeting.
Given the well-known role of location information in
social networking services, attempts to convert this information into mobility characteristics remain relatively
sparse. The most important foundation was provided by
Cheng et al. (2011), who analyzed different aspects of
mobility based on Twitter check-ins, which were at that
point dominated by the feed from another location-sharing
261
service – Foursquare. The study was extensive in scope
but was limited by the availability of temporal data. There
were other studies, such as that by Cho, Myers, and
Leskovec (2011), who used data from the Gowalla and
Brightkite services to model the influence of human mobility on social ties, or by Noulas et al. (2012) who focused
on intra-urban mobility derived from Foursquare
check-ins.
In this article, we present a global study of mobility
based on the analysis of Twitter data and the mobility
characteristics of different nations. We also seek to discover spatial patterns and clusters of regional mobility.
Finally, we attempt to validate the representativeness of
geo-located Twitter as a global source for mobility data.
The article is organized as follows. First, we describe the
dataset and illustrate a method to assign users to a country
of residence, hence enabling the determination of home
users and foreign visitors. Next, we present and compare
mobility profiles of various countries, as well as the temporal patterns of inflow and outflow dynamics. Further, we
explore country-to-country network of travels and delineate global regions of mobility. Finally, we validate the
results in two ways: (i) through a comparison of the
Twitter data with worldwide tourism statistics and (ii)
with commonly used models of human mobility.
Data pre-processing
Our study relies on one full year of geo-located tweets1
that were posted by users all over the world between
January 1, 2012, and December 31, 2012. The database
consists of 944M records generated by a total of 13M
users. The stream was gathered through the Twitter
Streaming API.2 Although the service sets a limit on
how much data can be accessed to less than 1% of the
total Twitter stream, the total geo-located content was
found not to exceed this restriction (Morstatter et al.
2013). Therefore, we believe that we successfully collected a complete picture of global geo-located Twitter
activity in 2012.
The database had to be cleaned before the analysis to
prevent the contamination of mobility statistics by errors
and artificial tweeting noise. First, we examined all the
consecutive locations of a single user and excluded those
that implied a user relocating with a speed over 1000 km/
h, i.e., faster than a passenger plane. Further, we filtered
out such activities on Twitter as web advertising (e.g.,
tweetmyjob), web gaming (e.g., map-game), or web
reporting (e.g., sandaysoft). Those services can generate
significant volumes of data, which do not reflect human
physical presence in either the reported place or time. To
correct for this artificial tweeting noise, we checked the
popularity of a message’s source, assuming that those with
only few users can probably be classified as an artificial
activity. As the threshold, we used a cumulative popularity
262
B. Hawelka et al.
territories, with the number of users greatly varying
among different countries. The unquestionable leader is
USA with over 3.8M users, followed by the UK,
Indonesia, Brazil, Japan, and Spain with over 500K
users each. There are also countries and territories with
only few or no Twitter users assigned.
To evaluate the representativeness of Twitter in a given
area, we used a more illustrative metric, the penetration
rate, defined as the ratio between the number of Twitter
users and the population of a country. As expected, this
rate does not distribute uniformly across the globe and
scales superlinearly with the level of a country’s economic
development (expressed as GDP per capita) (Figure 2A
and B). Although this property has already been described
by Mocanu et al. (2013) and others, we did observe that
the goodness of a power law fit increased when considering penetration of only residents rather than all Twitter
users appearing in a country. In the analysis, we exclude
all countries with a penetration rate below 0.05% (we also
exclude countries with less than 10,000 resident users).
Definition of users’ countries of residence
An essential first step in our cross-country mobility analysis was an explicit assignment of each user to a country
of residence. This made our work different from most of
the other Twitter studies that usually did not attempt to
uncover users’ origin and characterized a study area using
only the total volume of tweets observed in this area (e.g.,
Mocanu et al. 2013). While for certain research problems
this approach is suitable, from the perspective of a global
mobility study, the differentiation between residents and
visitors is crucial. It enables a clear definition of the origin
and destination of travels and reveals which nation is
traveling where and when. Taking advantage of the history
of tweeting records of each user, we defined country of
residence as the country where the user had issued most of
the tweets. Once the country of residence was identified,
the user’s activity in any other country of the world was
considered as traveling behavior, and the user was counted
as a visitor to that country.
We use the country definitions provided by the global
administrative areas spatial database, which divides the
world into 253 territories (Global Administrative Areas
2012). Twitter “residents” were identified in 243 of these
A
Twitter penetration
rate
up to 0.0002
0.0003 – 0.002
Mobility profiles of countries
Human mobility can be analyzed at different levels of
granularity. In this study, we considered a user as being
“mobile” if over the whole year the user had been tweeting from at least one country other than her or his country
of residence. This applied to 1M users or around 8% of all
those who used geo-located Twitter in 2012. Figure 3
shows the percentage of mobile users per country and
log Twitter penetration index
Downloaded by [] at 07:52 09 March 2016
among 95% of users, constructing the ranking separately
for each country. All tweeting sources falling below the
threshold were discarded from further analysis. In total,
the refinement procedure preserved 98% of users and 95%
of tweets from the initial database.
0.0021 – 0.006
0.0061 – 0.01
0.0101 – 0.0667
–2
B
10
–3
10
–4
10
–5
10
1000
10,000
100,000
log GDP per capita
Figure 2. Twitter penetration rate across the countries of the world (A). Spatial distribution of the index. (B) Superlinear scaling of the
penetration rate with per capita GDP of a country. R2 coefficient equals 0.70.
Figure 3.
Countries with the highest rates of users’ travel activity.
Cartography and Geographic Information Science
Downloaded by [] at 07:52 09 March 2016
the (geo-located) Twitter penetration in that country in
2012. Most of the top mobile countries, e.g., Belgium,
Austria, were characterized by only moderate levels of
Twitter adoption. On the other hand, users of geo-located
Twitter from the USA, the country with the highest penetration rate, revealed a surprisingly small tendency to
travel. The only two countries with high mobility and
penetration rates were Singapore and Kuwait. In general,
while an increased popularity of Twitter can be treated as a
sign of a more active society, it did not immediately imply
higher mobility of its users.
Next, we examined how spatially spread or concentrated the mobility of users is in a certain nation. This was
captured through an average radius of gyration of the
users. The radius of gyration measures the spread of
user’s locations around her or his usual location. Here,
we defined a usual location as the center of mass rather
than a home location, as the latter was defined too broadly,
at the level of a particular country. For each user, the
radius of gyration was computed thus:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
1X
acm j2
ai rg ¼
j
n i¼1
(1)
where n is the number of tweeting locations, āi represents
the location of a particular tweet (a pair of x–y coordinates), and ācm is a user’s center of mass. Low values of
the radius indicate a tendency to travel locally, while
higher values indicate more long-distance travels. The
average values computed for users from different countries are shown in Figure 4. At first glance, it is obvious
that the geographical location of a country played an
important role. Isolated countries such as New Zealand
or Australia had an average radius of gyration of over
700 km. There was also a positive correlation between
the average distance travelled by residents of a country
and the mobility rate of its Twitter population, as well as
the number of visited countries (Figure 4A and B). This
notwithstanding, the drivers for increased mobility were
263
invariably linked to the economic prosperity of a country,
as all received rankings were led by highly developed
countries.
The mobility profile of each country can be analyzed
from two perspectives, as being either the origin or the
destination of international travel. By building the directional country-to-country network of human travels, we
were able to quantify both the inflow and outflow of
visitors. Figure 5 shows the results of country-specific
analyses based on Twitter users and on the estimated
total number of travelers. Figure 5A shows the number
of Twitter users residing in a country and traveling to
another and Figure 5B shows the number of users visiting
this particular country. Figure 5C and D present the number of Twitter travelers normalized by the Twitter penetration rate in the user’s home country. In the case of inflow
of travelers, both the raw number of Twitter users and the
estimated population of visitors point to USA, UK, Spain,
and France to be the most visited countries. The nationality of outgoing travelers seemed highly influenced by
Twitter’s penetration rate, with the biggest groups coming
from countries of high Twitter popularity. Low penetration
indices lead to an overestimation of the actual volume of
travel, which may explain the high values estimated for
Russia or Germany. Figure 5E presents the yearly ratio
between the estimated inflow and outflow of travelers,
revealing which countries were the origin or destination
of international travel.
Temporal patterns of mobility
Human mobility is subject to temporal variations. In order
to uncover patterns occurring at the global level, as well as
at the country level, we measured how many Twitter users
were active outside of their country of residence for each
day of 2012. The first pattern that emerged from the
analysis was the weekly scheme of check-ins abroad
(Figure 6). The tendency of increased mobility over weekends seemed to be universal across the globe. Moreover,
there were two obvious seasons of higher mobility: the
Figure 4. Average radius of gyration of users from different countries compared to (A) percentage of mobile Twitter users and
(B) number of countries visited.
264
B. Hawelka et al.
Downloaded by [] at 07:52 09 March 2016
Figure 5. Number of visitors coming from or arriving in a country. (A and B) Number of Twitter travelers, (C and D) estimated total
number of travelers (number of Twitter travelers normalized by the Twitter penetration rate in the country of origin of the visitor), and
(E) the yearly ratio between the estimated inflow and outflow of travelers.
Figure 6.
Global temporal pattern of abroad travels by Twitter users.
summer months of July and August and the end of the
year, connected to Christmas and New Year’s Eve
holidays.
When specific countries were considered, we discovered a variety of deviations from the aforementioned global pattern. Several of them were easy to interpret and
were shared by more than one country. For instance, there
was a substantial group of European nations with the
biggest peak over one of the summer months and a few
smaller ones, most probably connected to extended weekends, e.g., at the beginning of May (examples are shown
in Figure 7A). Another group exhibited a similar pattern;
however, the summer mobility increase for this group
stretched between June and September (Figure 7B). An
interesting example of how the mobility behavior is influenced by the social and cultural norms of a country was
observed in a group of Arabic countries (Figure 7C). The
period of Ramadan corresponded to a major decrease in
the amount of travel abroad, while the time of the Mecca
pilgrimage at the end of October was marked by a sharp
peak. In all cases, increased international mobility corresponded with the end of the year.
The temporal variations of the inflow of visitors were
much more stable than those of the outflow patterns.
Visual inspection of those patterns identified three main
groups. The first group included countries without any
specific seasonality of international arrivals (with the
exception of the end of the year). The second group
covered popular summer destinations such as Spain,
Italy, Croatia, or Greece (Figure 8), with a significant
increase in arrivals over the months of July and August.
Finally, in the third group, we included countries where
increased international arrivals were connected to special
events such as Euro 2012 in Poland or the 2012 Olympics
in the UK.
Country-to-country network and partitioning
Next, we analyzed the topology of the country-to-country
mobility network created by travelers within the Twitter
community. As it has already been proven by many other
studies, partitioning of a raw network of human communication interactions, e.g., based on mobile phone data
(Blondel 2011; Blondel, Krings, and Thomas 2010; Ratti
et al. 2010; Sobolevsky, Campari, et al. 2013; Sobolevsky,
Szell, et al. 2013), as well as partitioning of human mobility (Amini et al. 2013; Kang et al. 2013), can lead to the
delineation of spatially cohesive communities, aligning
surprisingly well with the existing socioeconomic borders
of the underlying geographies. Our aim was to test if this
finding holds true for the Twitter-based mobility network,
Cartography and Geographic Information Science
265
Downloaded by [] at 07:52 09 March 2016
Figure 7. Normalized temporal patterns of mobility, by country of origin. The values for each country are scaled between 0 and 100%
of the maximum daily number of travelers being abroad during 2012.
Figure 8. Destinations of tourist activity with increased inflow of international Twitter users over summer. The values for each country
are scaled between 0 and 100% of the maximum daily number of international visitors during 2012.
and if so, which distinctive mobility clusters emerge in
different parts of the world.
Taking advantage of our methodology of assigning
users to their country of residence and focusing on mobile
users, we built a worldwide country-to-country network of
tweet flows. Each country was considered as a node, and
the edges of the network were weighted with the number of
Twitter users traveling between a pair of nodes. The network was directional, as the connections were built from
the country of residence to countries where a user appeared
as a visitor. To deal with the sparseness of the network and
different levels of Twitter representativeness, we filtered out
all countries with the outgoing Twitter population smaller
than 500 users and countries where the Twitter penetration
was below 0.05%. The flows were normalized by the
Twitter penetration rate in the country of a user’s origin
in order to estimate the real mobility flux rather than just a
number of Twitter users. The top 30 flows between different countries are presented in Figure 9.
The network partitioning procedure was based on the
well-known modularity optimization approach (Newman
2006) and uses a highly efficient optimization algorithm
recently proposed by Sobolevsky, Campari, et al. (2013).
Figure 9. Top 30 country-to-country estimated flows of visitors. Colors of the ribbons correspond to the destination of a trip;
the country of origin is marked with a thin stripe at the end of a
ribbon (visualization method based on Krzywinski et al. 2009).
Downloaded by [] at 07:52 09 March 2016
266
B. Hawelka et al.
The procedure assesses the relative strength of particular
links versus the estimations of the homogenous null
model. It optimizes the overall modularity score of a network partitioning, which quantifies the strength of
intracluster connections (in the “ideal” partitioning case
they should be as strong as possible) and the weakness of
outer ties (should be as weak as possible).
After obtaining the initial split of the network, the
partitioning procedure was applied iteratively to the subnetworks inside each community, much as Sobolevsky, Szell,
et al. (2013) did. As a result, the Twitter network was split
into mobility clusters on three hierarchical levels, each
level being a subpartitioning of the previous one. The initial
level (Figure 10A) uncovered four groups of countries that
reflected the continental division of the world. In this sense,
travel connections between North and South America were
stronger than that between America and Europe, while the
Europeans traveled more within Europe and to Asia than to
the other continents. Further subdivisions followed the
same type of logic, the clusters tending to be spatially
connected and well aligned with common socio-geographical regions. For instance, on level 2, we observed a split
into Western, Central, Eastern, and Northern Europe
(Figure 10B) and on level 3, Central Europe was further
divided into the more continental northern part and the
Balkans (Table 1). Received mobility clusters intimate
that people tend to travel more to close-by destinations
rather than further afield. Furthermore, the finding that
clusters were spatially continuous and reflected common
world regions is in line with the findings of previous
studies with mobile phone data (Blondel 2011; Blondel,
Krings, and Thomas 2010; Ratti et al. 2010; Sobolevsky,
Szell, et al. 2013). The finding also extends the validity of
network-based community detection from a country to
global scale. Additionally, we see that the partitioning of
mobility networks follows the same as human communication networks and might thus be used for regional delineation purposes.
Validation of the results
For global mobility, it is difficult to find a bias-free human
mobility dataset that would enable direct validation of
results obtained with Twitter. An interesting comparison
could be made, for instance, with the register of flight
connections, although this could be hampered by the
possibility of confounding a segment of an indirect travel
for direct travel from one’s home country to an intended
destination. In practice, such a comparison is also
obviously prevented by the difficulty of accessing such
data. In this study, we relied on tourism statistics provided
by the World Economic Forum (WEF 2013) at the country
level. We used two of those statistics: international tourist
arrivals (thousands, 2011) and international tourism
Figure 10. Mobility regions uncovered by the partitioning of the country-to-country network of Twitter user flows. Regions distinguished at the first (A) and second (B) level of partitioning. Gray color indicates no data.
Cartography and Geographic Information Science
Table 1.
Level 1
1
Countries assigned to different regions of mobility.
Level 2
Level 3
Assigned countries
1
2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
Bahamas, Canada, Dominican Republic, Jamaica, Mexico, Puerto Rico, USA
Colombia, Ecuador, Panama, Trinidad and Tobago, Venezuela
Costa Rica, El Salvador, Guatemala, Honduras
Bolivia, Chile, Peru
Argentina, Brazil, Paraguay, Uruguay
France, Ireland, Malta, Martinique, Morocco, Portugal, Spain, Tunisia, UK
Belgium, Germany, Iceland, Italy, Luxembourg, Netherlands, Switzerland
Denmark, Norway, Sweden
Austria, Czech Republic, Hungary, Poland, Romania, Slovakia
Bosnia and Herzegovina, Croatia, Kosovo, Macedonia, Serbia, Slovenia
Azerbaijan, Bulgaria, Cyprus, Greece, Israel, Kazakhstan, Latvia, Lithuania, Russia, Ukraine
Belarus, Estonia, Finland, Turkey
Ghana, Nigeria
Kenya, South Africa
Bahrain, Egypt, Jordan, Kuwait, Lebanon, Saudi Arabia
Oman, Qatar, United Arab Emirates
Maldives, Sri Lanka
Japan, South Korea, Taiwan
Philippines
Brunei, Indonesia, Malaysia, Singapore
Cambodia, Thailand, Vietnam
Australia, New Zealand
3
2
4
5
6
7
3
8
9
10
Downloaded by [] at 07:52 09 March 2016
4
267
11
12
13
14
receipts (US$, millions, 2011) and compared them to
arrivals estimated on the basis of the Twitter data
(Figures 11A and 10B). In both cases, we found a strong
linear correlation (respectively with the R2 of 0.69 and
0.88), which confirms the validity of the estimated mobility figures.
We further validated the results indirectly by demonstrating that mobility measures derived from Twitter activity exhibit similar statistical properties as those obtained
using other datasets. First, we computed the distance
between each pair of consecutive user locations (tweets)
and plotted the frequencies of computed displacement on a
Figure 11. International arrivals estimated with Twitter data versus the arrivals (A) and nominal value of tourist receipts (expenditures
by international inbound visitors, B) provided by WEF (2013). R2 statistic equals 0.69 and 0.88, respectively.
Figure 12.
Probability of displacement (A) and frequency of the radius of gyration (B).
268
B. Hawelka et al.
log-log scale. Similarly to other studies, we found that the
distribution is well approximated by the power law
(Figure 12A):
Downloaded by [] at 07:52 09 March 2016
PðΔrÞ ¼ Δrβ
(2)
where Δr is a displacement of certain length and β = 1.62.
Importantly, the received exponent stays in the similar
range as the exponents obtained with other mobility datasets
such as mobile phone data (González, Hidalgo, and Barabási
2008, β = 1.75), bank note dispersal (Brockmann, Hufnagel,
and Geisel 2006, β = 1.59), and Foursquare check-ins
(Cheng et al. 2011, β = 1.88). We also plotted the frequency
of previously computed users’ radiuses of gyration
(Figure 12B). As expected, it also followed a power law
with an exponent of 1.25.
Given the limited access to global mobility data suitable for a direct comparison with Twitter-based human
travels, we further tested our data against a commonly
accepted mobility model – the classic gravity approach –
as yet another way of indirect verification of uncovered
patterns. Many studies proved the gravity law to provide a
good basis for modeling the intensity of interactions
between locations, depending on their weights and distance, in the context of not only mobility (e.g., Balcan
et al. 2009; Jung, Wang, and Stanley 2008; Zipf 1946) but
also human interaction networks (Expert et al. 2011;
Krings et al. 2009). We made the assumption that if the
Twitter data were to be considered suitable for a description of human mobility, they should follow a similar law
and exhibit similar distance dependence as those found for
railway connections, airline traffic, mobile telecom
records, and other data. We also tested how much the
gravity law holds on a global scale, especially in times
when the influence of distance as a barrier to mobility is
often considered to decrease in importance. We used the
gravity model in the form:
Fij ¼ A
pαi pjβ
rijγ
(3)
where Fij represents a flow of people between a pair of
countries, pi is the population in the country of origin, pj
the population of the destination country, rij is the distance
measured between the capitals, and A is a constant. To
compensate for the limitation of our definition of countryto-country distance, we restricted the connections to the
countries that are at least 100 km apart. The model was
fitted to the two versions of our network. First, the flows
were defined as a raw number of Twitter users traveling
between two countries, and the population of each country
was equal to the number of Twitter residents (Figure 13A).
The exponents were α = 0.81, β = 0.63, and γ = 1.02 and
the R2 coefficient was 0.79. The second variation of the
model (Figure 13B) used flows estimated based on the
Twitter penetration rate (again with the threshold of
0.05%) and population provided by the Central
Intelligence Agency (2012). The exponents found were
fairly similar to the previously found exponents –
α = 0.89, β = 0.69, and γ = 1.1 – but the R2 coefficient
(0.71) was slightly lower.
The population exponents we obtained suggest that a
country’s size influences the growth of human flow sublinearly, in terms of both origin and destination of travel,
but the influence of the population in a country of origin is
bigger. This relationship can be explained with two conjectures. On the one hand, it is plausible that residents of a
country do not take part in mobility in an equal manner;
rather it is a domain of the most active residents. On the
other hand, it could be that visitors were never attracted by
the whole country but only by certain places within the
country. The number of active users and attractive places
in countries may on average grow slower than the countries’ total population.
The gamma exponents suggest, as expected, a
decrease of interaction intensity with distance
(Figure 13A and B), however at a slower rate than often
pre-assumed r2 decay relationship, e.g., by Jung, Wang,
and Stanley (2008) or Krings et al. (2009). The difference
in received decay relationship can be explained by the
scale of analysis. On a global scale, where most of the
trips are happening by air, an increase in distance comes
with relatively smaller effort or cost than on a country or
Figure 13. Dependence of human flow (Fij) normalized with populations in countries of origin and destination (A piαpjβ) on the distance
in comparison to a distance decay function (r–γ) modeled with the gravity law. (A) Network defined based on raw Twitter flows.
(B) Network of total population flows estimated with the Twitter penetration rate in the country of origin.
Cartography and Geographic Information Science
local levels where most travel is by land. But even in the
world of this subjective “shrinkage” of distances, certain
level of dependency is preserved, possibly because of
social ties, which remain stronger on a local than global
scale (Takhteyev, Gruzd, and Wellman 2012). In other
words, people may simply have more reasons to travel
shorter distances, which also correlates well with the
results received during network partitioning.
Visually, both models seem to be well fitted (Figure
13A and B), with the slopes reflecting well the average
tendency in the observed data and the distant decay functions remaining within the standard errors across the distance range. The similarity of both models suggests that
Twitter data not only may provide a valid picture of
mobility of its direct users, but can further be used for
the estimation of real human flows.
Downloaded by [] at 07:52 09 March 2016
Conclusions
Geo-located Twitter is one of the first free and easily
available global data sources that store millions of objective, digital records of human activity in space and time.
In our study, we demonstrated that, despite the unequal
distribution over the different parts of the world and
possible bias toward a certain part of the population, in
many cases geo-located Twitter can and should be considered as a valuable proxy for human mobility, especially
at the level of country-to-country flows. Our approach
proposed to capture nation-specific mobility by assigning
users to their country of residence. As a result, we were
able to compare mobility profiles of countries, considering
each country as both a potential origin and a destination of
international travels. The results showed that increased
mobility (measured in terms of the probability of travel,
diversity of destinations, and geographical spread of travels) is characteristic of West European and other developed countries. Traveling distance was additionally
affected by the geographic isolation of a country, such as
Australia or New Zealand. Through the analysis of temporal patterns, we found a globally universal season of
increased mobility at the end of the year. Although the
summer mobility was increased for a wide range of countries, it varied in terms of intensity and duration, and in
some cases there was no increase at all. Additionally, we
discovered patterns driven either by cultural conditioning
or by special events occurring in a country. In many cases,
the results well confirmed logical expectations, which we
treat as an indicator of the legitimacy of Twitter as a global
and objective register of human mobility.
Furthermore, we demonstrated that the communities
detected using a Twitter mobility network formed spatially
cohesive regions reflecting the regional division of the
world. This finding is important from several standpoints.
First, it is in agreement with the results obtained by
Blondel, Krings, and Thomas (2010), Ratti et al. (2010),
269
and Sobolevsky, Szell, et al. (2013) who based their
research on networks of mobile phone interactions.
Second, it expands the spatial validity of the community
detection approach from a previously examined country
scale to the global scale. Third, it shows that even in the
era of globalization and seeming decrease of the influence
of distance, people still tend to travel locally, visiting
neighboring countries more often than those further away.
Further, we validated, to a certain extent, geo-located
Twitter as a proxy of global mobility behavior. We demonstrated that the number of visitors estimated for different
countries based on Twitter data is in line with the official
statistics on international tourism. The correlation (R2
around 0.7) shows a fairly good correspondence, given
the wider scope of mobility captured through Twitter and a
significantly different method of data acquisition. Further,
we confirmed that Twitter data exhibit similar statistical
properties as other mobility datasets. For instance, measures such as radius of gyration and probability of displacement are well estimated with the power law distribution
(similarly as in Cheng et al. 2011; Brockmann, Hufnagel,
and Geisel 2006), while the network of estimated flows of
travelers can be described fairly accurately using the classic model of a mobility – the gravity model.
We believe the analysis presented in this article proves
the potential of geo-located Twitter as an objective, freely
accessible source of data for global mobility studies.
Further research will focus on exploring the applicability
of Twitter activity for finer spatial scales.
Acknowledgments
This research was funded by the Austrian Science Fund (FWF)
through the Doctoral College GIScience (DK W 1237-N23),
Department of Geoinformatics – Z_GIS, University of Salzburg,
Austria. We thank Sebastian Grauwin and Karolina Stanislawska
for their support. We further thank the MIT SMART Program, the
Center for Complex Engineering Systems (CCES) at KACST and
MIT CCES program, the National Science Foundation, the MIT
Portugal Program, the AT&T Foundation, Audi Volkswagen,
BBVA, The Coca Cola Company, Ericsson, Expo 2015, Ferrovial,
GE, and all the members of the MIT Senseable City Lab
Consortium for supporting the research.
Notes
1.
2.
By geo-located, we mean the messages with explicit geographic coordinates attached to each message.
https://dev.twitter.com/docs/streaming-apis
References
Amini, A., K. Kung, C. Kang, S. Sobolevsky, and C. Ratti. 2013.
“The Differing Tribal and Infrastructural Influences on
Mobility in Developing and Industrialized Regions.” In
Proceedings of 3rd International Conference on the
Analysis of Mobile Phone Data (NetMob 2013), 330–339.
Cambridge, MA: MIT.
Downloaded by [] at 07:52 09 March 2016
270
B. Hawelka et al.
Bajardi, P., C. Poletto, J. J. Ramasco, M. Tizzoni, V. Colizza, and
A. Vespignani. 2011. “Human Mobility Networks, Travel
Restrictions, and the Global Spread of 2009 H1N1
Pandemic.” PLoS ONE 6 (1): e16591. doi:10.1371/journal.
pone.0016591.
Balcan, D., V. Colizza, B. Gonçalves, H. Hu, J. J. Ramasco, and
A. Vespignani. 2009. “Multiscale Mobility Networks and the
Spatial Spreading of Infectious Diseases.” Proceedings of the
National Academy of Sciences 106 (51): 21484–21489.
doi:10.1073/pnas.0906910106.
Barrat, A., M. Barthélemy, R. Pastor-Satorras, and A. Vespignani.
2004. “The Architecture of Complex Weighted Networks.”
Proceedings of the National Academy of Sciences of the
United States of America 101 (11): 3747–3752. doi:10.1073/
pnas.0400087101.
Blondel, V. 2011. “Voice on the Border: Do Cellphones Redraw
the Maps?” Open Access publications from Université
Catholique de Louvain info: hdl:2078.1/108639. Luvain-laNeuve: Université Catholique de Louvain.
Blondel, V., G. Krings, and I. Thomas. 2010. “Regions and
Borders of Mobile Telephony in Belgium and in the
Brussels Metropolitan Zone.” Brussels Studies 42 (4): 1–12.
Bollen, J., A. Pepe, and H. Mao. 2011. “Modeling Public Mood and
Emotion: Twitter Sentiment and Socio-economic Phenomena.”
In Proceedings of the Fifth International AAAI Conference on
Weblogs and Social Media, 450–453. http://www.aaai.org/ocs/
index.php/ICWSM/ICWSM11/paper/viewFile/2826/3237
Brockmann, D., L. Hufnagel, and T. Geisel. 2006. “The Scaling
Laws of Human Travel.” Nature 439 (7075): 462–465.
doi:10.1038/nature04292.
Calabrese, F., M. Diao, G. Di Lorenzo, J. Ferreira Jr., and C.
Ratti. 2013. “Understanding Individual Mobility Patterns
from Urban Sensing Data: A Mobile Phone Trace
Example.” Transportation Research Part C: Emerging
Technologies 26: 301–313. doi:10.1016/j.trc.2012.09.009.
Calabrese, F., F. C. Pereira, G. Di Lorenzo, L. Liu, and C. Ratti.
2010. “The Geography of Taste: Analyzing Cell-Phone
Mobility and Social Events.” In Pervasive Computing. Lecture
Notes in Computer Science, Vol. 6030, 22–37. Berlin: Springer.
Castles, S., and M. J. Miller. 1998. The Age of Migration:
International Population Movements in the Modern World.
New York: Guildford Press.
Central Intelligence Agency (CIA). 2012. The World Factbook.
Accessed September 27, 2013. https://www.cia.gov/library/
publications/the-world-factbook/
Cheng, Z., J. Caverlee, K. Lee, and D. Z. Sui. 2011. “Exploring
Millions of Footprints in Location Sharing Services.” In
Proceedings of the Fifth International AAAI Conference on
Weblogs and Social Media, Barcelona, June 17–21, 81–88.
Menolo Park, CA: AAAI Press.
Cho, E., S. A. Myers, and J. Leskovec. 2011. “Friendship and
Mobility: User Movement in Location-based Social
Networks.” In Proceedings of the 17th ACM SIGKDD
International Conference on Knowledge Discovery and
Data Mining, 1082–1090. New York: ACM.
Expert, P., T. S. Evans, V. D. Blondel, and R. Lambiotte. 2011.
“Uncovering Space-Independent Communities in Spatial
Networks.” Proceedings of the National Academy of Sciences
108 (19): 7663–7668. doi:10.1073/pnas.1018962108.
Frank, M. R., L. Mitchell, P. S. Dodds, and C. M. Danforth.
2013. “Happiness and the Patterns of Life: A Study of
Geolocated Tweets.” Scientific Reports 3: 1–9. doi:10.1038/
srep02625.
Frias-Martinez, V., V. Soto, H. Hohwald, and E. Frias-Martinez.
2012. “Characterizing Urban Landscapes Using Geolocated
Tweets.” In Privacy, Security, Risk and Trust (PASSSAT),
2012 International Conference on Social Computing
(SocialCom), 239–248. Amsterdam: IEEE.
Gesenhues, A. 2013. “Study: Mobile Users & Older Generations
are Driving Social Media Growth around the World.”
Marketing Land Publication Service. Third Doors Media.
Accessed September 27. http://marketingland.com/studysocial-network-growth-across-the-globe-driven-by-mobileusers-older-generations-41982
Ghosh, D., and R. Guha. 2013. “What are we ‘Tweeting’ about
Obesity? Mapping Tweets with Topic Modeling and
Geographic Information System.” Cartography and
Geographic Information Science 40 (2): 90–102. doi:10.1080/
15230406.2013.776210.
Global Administrative Areas. 2012. GADM Database of Global
Administrative Areas. Accessed August 15, 2013. http://
www.gadm.org/
Golder, S. A., and M. W. Macy. 2011. “Diurnal and Seasonal Mood
Vary with Work, Sleep, and Daylength across Diverse Cultures.”
Science 333 (6051): 1878–1881. doi:10.1126/science.1202775.
González, M. C., C. A. Hidalgo, and A. L. Barabási. 2008.
“Understanding Individual Human Mobility Patterns.”
Nature 453 (7196): 779–782. doi:10.1038/nature06958.
Greenwood, M. J. 1985. “Human Migration: Theory, Models,
and Empirical Studies.” Journal of Regional Science 25 (4):
521–544. doi:10.1111/j.1467-9787.1985.tb00321.x.
Huberman, B. A., D. M. Romero, and F. Wu. 2008. “Social
Networks That Matter: Twitter Under the Microscope.”
SSRN Scholarly Paper ID 1313405. http://papers.ssrn.com/
abstract=1313405
Java, A., X. Song, T. Finin, and B. Tseng. 2007. “Why We
Twitter: Understanding Microblogging Usage and
Communities.” In Proceedings of the 9th WebKDD and 1st
SNA-KDD 2007 Workshop on Web Mining and Social
Network Analysis, 56–65. New York: ACM.
Jung, W.-S., F. Wang, and H. E. Stanley. 2008. “Gravity Model
in the Korean Highway.” Europhysics Letters 81 (4): 48005.
doi:10.1209/0295-5075/81/48005.
Kamath, K. Y., J. Caverlee, Z. Cheng, and D. Z. Sui. 2012.
“Spatial Influence vs. Community Influence: Modeling the
Global Spread of Social Media.” In Proceedings of the 21st
ACM International Conference on Information and
Knowledge Management, 962–971. Maui, HI: ACM.
Kang, C., X. Ma, D. Tong, and Y. Liu. 2012. “Intra-Urban Human
Mobility Patterns: An Urban Morphology Perspective.”
Physica A: Statistical Mechanics and Its Applications 391 (4):
1702–1717. doi:10.1016/j.physa.2011.11.005.
Kang, C., S. Sobolevsky, Y. Liu, and C. Ratti. 2013. “Exploring
Human Movements in Singapore: A Comparative Analysis
Based on Mobile Phone and Taxicab Usages.” In
UrbComp’13 Proceedings of the 3rd ACM SIGKDD
International Workshop on Urban Computing 330–339.
Chicago, IL: ACM.
Krings, G., F. Calabrese, C. Ratti, and V. D. Blondel. 2009. “Urban
Gravity: A Model for Inter-City Telecommunication Flows.”
Journal of Statistical Mechanics: Theory and Experiment 2009
(7): L07003. doi:10.1088/1742-5468/2009/07/L07003.
Krzywinski, M., J. Schein, I. Birol, J. Connors, R. Gascoyne, D.
Horsman, S. J. Jones, and M. A. Marra. 2009. “Circos: An
Information Aesthetic for Comparative Genomics.” Genome
Research 19 (9): 1639–1645. doi:10.1101/gr.092759.109.
Kwak, H., C. Lee, H. Park, and S. Moon. 2010. “What Is Twitter,
a Social Network or a News Media?” In Proceedings of the
19th International Conference on World Wide Web, 591–600.
Raleigh, NC: ACM.
Downloaded by [] at 07:52 09 March 2016
Cartography and Geographic Information Science
Leetaru, K., S. Wang, A. Padmanabhan, and E. Shook. 2013.
“Mapping the Global Twitter Heartbeat: The Geography of
Twitter.” First Monday 18 (5). doi:10.5210/fm.v18i5.4366.
MacEachren, A. M., A. C. Robinson, A. Jaiswal, S. Pezanowski, A.
Savelyev, J. Blanford, and P. Mitra. 2011. “Geo-Twitter Analytics:
Applications in Crisis Management.” Proceedings, 25th
International Cartographic Conference, Paris, July 3–8.
Miguéns, J. I. L., and J. F. F. Mendes. 2008. “Travel and
Tourism: Into a Complex Network.” Physica A: Statistical
Mechanics and its Applications 387 (12): 2963–2971.
doi:10.1016/j.physa.2008.01.058.
Mitchell, L., K. D. Harris, M. R. Frank, P. S. Dodds, and C. M.
Danforth. 2013. “The Geography of Happiness: Connecting
Twitter Sentiment and Expression, Demographics, and
Objective Characteristics of Place.” ArXiv e-print arXiv:
1302.3299.
Mocanu, D., A. Baronchelli, N. Perra, B. Gonçalves, Q. Zhang,
and A. Vespignani. 2013. “The Twitter of Babel: Mapping
World Languages through Microblogging Platforms.” PLoS
ONE 8 (4): e61981. doi:10.1371/journal.pone.0061981.
Morstatter, F., J. Pfeffer, H. Liu, and K. M. Carley. 2013. “Is the
Sample Good Enough? Comparing Data from Twitter’s
Streaming Api with Twitter’s Firehose.” In Proceedings of
ICWSM. Cambridge, MA: AAAI Press.
Newman, M. E. J. 2006. “Modularity and Community Structure in
Networks.” Proceedings of the National Academy of Sciences
103 (23): 8577–8582. doi:10.1073/pnas.0601602103.
Noulas, A., S. Scellato, R. Lambiotte, M. Pontil, and C. Mascolo.
2012. “A Tale of Many Cities: Universal Patterns in Human
Urban Mobility.” PLoS ONE 7 (5): e37027. doi:10.1371/
journal.pone.0037027.
Pak, A., and P. Paroubek. 2010. “Twitter as a Corpus for
Sentiment Analysis and Opinion Mining.” In Proceedings
of LREC, 436–439. Valletta: European Language Resources
Association (ELRA).
Ratti, C., S. Sobolevsky, F. Calabrese, C. Andris, J. Reades, M.
Martino, R. Claxton, and S. H. Strogatz. 2010. “Redrawing
the Map of Great Britain from a Network of Human
Interactions.” PLoS ONE 5 (12): e14248. doi:10.1371/jour
nal.pone.0014248.
Sagl, G., B. Resch, B. Hawelka, and E. Beinat. 2012. “From
Social Sensor Data to Collective Human Behaviour
Patterns: Analysing and Visualising Spatio-Temporal
271
Dynamics in Urban Environments.” In Proceedings of the
GI-Forum 2012: Geovisualization, Society and Learning,
54–63. Salzburg: Wichmann.
Sakaki, T., M. Okazaki, and Y. Matsuo. 2010. “Earthquake
Shakes Twitter Users: Real-time Event Detection by Social
Sensors.” In Proceedings of the 19th International
Conference on World Wide Web, 851–860. Raleigh, NC:
ACM.
Sassen, S. 1999. Globalization and its Discontents: Essays on the New
Mobility of People and Money. New York: The New Press.
Simini, F., M. C. González, A. Maritan, and A. L. Barabási. 2012.
“A Universal Model for Mobility and Migration Patterns.”
Nature 484 (7392): 96–100. doi:10.1038/nature10856.
Sobolevsky, S., R. Campari, A. Belyi, and C. Ratti. 2013. “A
General Optimization Technique for High Quality
Community Detection in Complex Networks.” ArXiv
e-print arXiv: 1308.3508.
Sobolevsky, S., M. Szell, R. Campari, T. Couronné, Z.
Smoreda, and C. Ratti. 2013. “Delineating Geographical
Regions with Networks of Human Interactions in an
Extensive Set of Countries.” PloS ONE 8 (12): e81707.
doi:10.1371/journal.pone.0081707.
Takhteyev, Y., A. Gruzd, and B. Wellman. 2012. “Geography of
Twitter Networks.” Social Networks 34 (1): 73–81.
doi:10.1016/j.socnet.2011.05.006.
Thom, D., H. Bosch, S. Koch, M. Worner, and T. Ertl. 2012.
“Spatiotemporal Anomaly Detection through Visual
Analysis of Geolocated Twitter Messages.” In Pacific
Visualization Symposium (PacificVis), 2012 IEEE, 41–48.
Songdo: IEEE.
Twitter Statistics. 2013. Statistic Brain Research Institute,
Publishing as Statistic Brain. Accessed September 27.
http://www.statisticbrain.com/twitter-statistics
Wakamiya, S., R. Lee, and K. Sumiya. 2011. “Urban Area
Characterization Based on Semantics of Crowd Activities
in Twitter.” GeoSpatial Semantics 108–123. doi:10.1007/
978-3-642-20630-6_7.
WEF (World Economic Forum). 2013. The Travel & Tourism
Competitiveness Report 2013. http://www.weforum.org/
reports/travel-tourism-competitiveness-report-2013
Zipf, G. K. 1946. “The P1P2/D Hypothesis: On the Intercity
Movement of Persons.” American Sociological Review 11:
677–686. doi:10.2307/2087063.