Content uploaded by Elena Daehnhardt
Author content
All content in this area was uploaded by Elena Daehnhardt on Dec 18, 2014
Content may be subject to copyright.
Cultural and Geolocation Aspects of Communication in Twitter
E. Daehnhardt1, Y. Jing2, N.K. Taylor1
School of Mathematical & Computer Sciences, Heriot-Watt University
School of Computing, London Metropolitan University
elena@daehnhardt.com, y.jing@londonmet.ac.uk, N.K.Taylor@hw.ac.uk
Abstract
Web applications exploit user information from so-
cial networks and online user activities to facilitate
interaction and create an enhanced user experience.
Due to privacy issues however, it might be difficult
to extract user data from social network, in particu-
lar location data. For instance, information on user
location depends on users’ agreement to share own ge-
ographic data. Instead of directly collecting personal
user information, we aim to infer user preferences by
detecting behavior patterns from publicly available
micro blogging content and users’ followers’ network.
With statistical and machine-learning methods, we
employ Twitter-specific features to predict country
origin of users on Twitter with an accuracy of more
than 90% for users from the most active countries.
We further investigate users’ interpersonal communi-
cation with their followers. Our findings reveal that
belonging to a particular cultural group is playing an
important role in increasing users responses to their
friends. The knowledge on user cultural origins thus
could provide a di↵erentiated state-of-the-art user ex-
perience in microblogs, for instance, in friend recom-
mendation scenario.
1 Introduction
Social networking sites such as Twitter microblogs
allow users to communicate with their online friends
and share information in real time. Some web and
mobile applications require information on user loca-
tion and origin to provide users with location-specific
information, like recommending places of interest, so-
cial events or online friends in close proximity. A user
profile containing user geographic locations, on which
recommendations can be provided may impose also
possible privacy threats. This is why a majority of
Twitter users avoid sharing their accurate locations
[1].
However, despite of user e↵orts to hide or obscure
one whereabouts, there are methods to identify user
origins based on content [2, 3, 4], social profile meta-
data [2, 5], and from other social networking sites, in
which users’ data is gathered [6]. Besides location-
related cues, microblogging activity patterns can re-
veal users from di↵erent origins, which could be used
for indirectly inferring origins [7].
Determining precise user locations is discussed in
previous works on a city-level granularity [8]. Striv-
ing to preserve user privacy, we abstract from mining
accurate location-specific information.
We limit ourselves to a country-level or a “cultural
group” comprising a group of countries with further
defined behavioral patterns. User similarities and dif-
ferences found in the profile meta-data, contact net-
works, content and microblogging behavior can be
employed as a proxy for finding user origins, whether
they are country or culture-related. Our contribu-
tions to social networking research are the following:
•Investigation into the predictive value of Twitter
user-related and social-network related features
in experiments for predicting user origins.
•Analysis of user communication preferences in
Twitter on the example of followers’ responses
to their friends.
In section 2, we outline related work in the scope
of Twitter location detection, cultural aspects playing
a role in adaptation and personalization. In section
3, we describe our research methodology and experi-
mental setup. Next, we outline and discuss the results
in user origin and response prediction experiments.
We also discuss benefits and limitations of the pro-
posed approach and possible areas of improvement.
We conclude with research findings, how our findings
could be used in the development of adaptive social
web applications.
2 Related Work
One of the main requirements for adaptation is user
location, which can be used to mine location-specific
content [9] and user interests [10]. Explicit user lo-
cation, however, is often missed or not accurate, and
only a very small part of openly-available microposts
are geographically tagged [1]. This is why location
detection out of social web is a pertinent research
topic.
A method to detect locations from content is the
usage of a gazetteer or toponym vocabulary com-
prised of location-specific terms. However, the ap-
plication of gazetteer for location information dis-
ambiguation in microblogs is challenging for simi-
larly named locations and the inherent difficulties
of mining information out of the microblogs. Mis-
spelling and usage of abbreviations is common due
to the short message limitations [11]. To improve
performance of location detection using GeoNames
gazetteer, [11] employ Support Vector Machines clas-
sifier using Twitter features extracted from tweet con-
tent and meta-data.
As assessed by human annotators, geo-coding ser-
vices Yahoo and Google applied to profile locations on
Twitter are not really trustful for a large proportion
of tweets [12]. This is not surprising since about 30%
of users do not provide an accurate location [1]. As
[2] pointed out, user location is influenced by tempo-
ral dynamics and requires a model update when used
as a feature in a location detecting classifier. Named-
entity recognition with Stanford NER and Open NLP
tools is investigated in [13], showing a considerable
performance when training on Twitter-specific con-
tent, which requires human involvement for annotat-
ing data. Location disambiguation of Tweets with
Stanford NER, gazetteer, heuristic rules was per-
formed with precision and recall of around 80%, and
is comparable with human annotations [3]. Authors
also suggest that representation of the tweets’ content
with help of the ontologies [3] might be useful in the
toponym disambiguation task.
The detection of the home country from Twitter
content was also investigated in [1] with machine
learning technique whilst analyzing tweets’ content,
disregarding other Twitter features and the extended
contacts’ network of the users. Using Naive Bayes
classification model working with term frequencies,
user country locations were inferred in about 73% of
cases [1]. Geographical topic models created based on
terms extracted out of content correlate with specific
geographic locations, but require an adjusted proba-
bility estimation with help of smoothing technique in
order to deal with term sparsity [14, 10, 4]. Location-
specific terms selection with Kernel Density Estima-
tion is investigated in [15], demonstrating robustness
of the approach when only a small set of terms is
employed.
[2] exploit location, user name, description and
time zone fields for creation of a location-detecting
classifier, finding country location with an accuracy of
92% for the best feature set analyzed. [8] created an
ensemble classifier for detecting users’ home location
based on words and hashtags extracted from tweets,
tweets’ frequency dynamics and gazetteer dictionary
of geographic place names. Their classification algo-
rithm enabled hierarchical location detection of time
zone, state and city name with recall figures of 0.78,
0.66 and 0.58 respectively. For improving location
detection outcomes for Flickr images, [5] use statisti-
cal inference, gazetteer and other features extracted
from the Flickr users’ meta-data and content.
In location disambiguation based on user-generated
content, adding content of contact users helped to
improve prediction performance [16]. Locations of
users sharing web links help to predict the link ori-
gins [17]. Locations of users can be predicted based
on their contact networks, also considering di↵erent
social platforms [6].
Observing Twitter connections within national and
international geographical boundaries, [18] discuss
the nations as defining cultural representations of
communities, in which people communicate di↵er-
ently due to their di↵erent interests and geo-political
status of their countries. There are patterns of
follower-friend relationships, in which US users are
usually followed by users from other countries, while
Japanese are usually followed mostly by Japanese
[18]. Interestingly, the majority of connections are
within national borders [19, 18]. In accord with
[18], 39% of connections are within 100 km distance.
English-speaking users follow also English-speaking
users in the majority of over 90%, while other lan-
guage users build connections with users of the same
language, in 60% of cases [18].
There are several research domains investigating
adaptation and personalization outcomes consider-
ing cultural backgrounds of users. In e-learning, [20]
found that the cultural context of a learner might
impose requirements on technology usage, selection
of media and the style of interaction between stu-
dents and instructors. The learning material presen-
tation methods impact the performance of students
from di↵erent cultural groups [21]. To improve the
experience of e-learners in e-learning environments,
it is paramount to distinguish between di↵erent cul-
tural preferences [20].
In a recommendation system context, [22] analy-
Table 1: Research Question and Hypothesis
RQ1 How can Twitter microblogs be exploited to infer user origin and locality?
H1.1 The information on country location derived from user tweets’ meta-data and respective meta-data of
the followers is not sufficient for providing cues on user origins for the majority of users (75% from our
dataset).
RQ2 Which Twitter-specific features could be used for inferring the country location of users?
H2.1 A users contact network can assist in improving country prediction.
H2.2 User microblogging patterns can assist in further improving locality prediction.
RQ3 Which friend features are important for predicting a user’s follower responses?
H3.1 User and follower’s country locations’ and language match are amongst the most important prediction
parameters for user responses.
H3.2 User’s influence is significant in predicting her follower responses.
ses recommendation prospects for urban planning in
accord with user cultural similarities on Foursquare
and user preferences. Cultural behavior di↵erences
in user activities on Twitter were investigated in [23],
suggesting to exploit such di↵erences for building re-
sponsive communication applications. Other possi-
ble applications include friendship recommendations
based on suggestions by the network [24] and a com-
munity detection approach based on user interactions
on Twitter [25]. Microblogging response prediction
with focus on English-speaking users was studied in
[26], which analyzed the importance of specific terms
occurred in tweets, previous response ratios and also
number of user and follower links in the social net-
work. The importance of social relationship between
Twitter users on their responses prediction revealed
in [27].
Overall, the above mentioned research points out
cultural behavior di↵erences of users online. It re-
mains however unclear whether knowledge on the user
cultural background would help in improving recom-
mendation performance, and whether such recom-
mendations would outperform country-specific rec-
ommendations. A further in-depth investigation into
follower and friend relationships on Twitter could
shed a light into location-specific cultural aspects
playing a role in online communication. For pre-
dicting user responses, we further evaluate the inclu-
sion of the geographic and culture-specific aspects of
the users involved. For this, we examined the user
and follower-related features enabling to identify user
country and culture-level origins, which we further
exploited in a user response (reply or retweet cases)
prediction experiment.
3 Methodology
The main aim of the study was to analyze user com-
munication in Twitter, in order to get insights into
friend-follower relationships for further reply predic-
tions. It seems reasonable to assume that user geo-
graphic locations, cultural origin and language might
play a role in the follower interests reflected in reply
or retweet messages.
We also analyzed the relative importance of other
Twitter-specific features. For instance, number of
user friends and followers, which ratio is often ex-
plained as an “Influence” on other users might help
us to detect interesting users to follow. Also, we as-
sess user-related and follower-related features for pre-
dicting a user whereabouts and the interest of the
followers. We present our research questions and hy-
pothesis in Table 1. Our research is based on the
following assumptions and definitions:
•cultural group is the group of individuals com-
prising one or several nations exhibiting similar
behavior and communication traits as previously
explored in the sociological studies such as Hofst-
ede, which application in information system re-
search is critically assessed in [28]. We are aware
that not all individuals are expected to exhibit
behavior strictly within an associated cultural or
country group. However, we employ the cultural
group notion to merely stereotype online users.
•Based on previous research [7], we employed the
Lewis model of cultures describing how persons
belonging to di↵erent cultural backgrounds dif-
fer in their interpersonal behavior. For instance,
Asian cultures, such as Japan or Vietnam, are
defined by Lewis in their culture dimension as
Reactive (RE), since they are generally consid-
ered to be courteous, accommodating and good
listeners. In addition, Lewis defines a Multi-
Active (MA) and Linear-Active (LA) cultural di-
mension. While persons described as MA focus
on interpersonal communication and are gener-
ally considered as emotional personalities, LA
(a) Languages (percentages) and Countries (ISO codes), the
Ver t i cal Ax i s S how s t he Num b e r of Use r s
(b) User Locations and Dimensions (blue: LA, yellow: RE,
red: MA)
Figure 1: Top Languages and Geographic Locations
persons focus on working with facts and plan-
ning activities [29].
•Therefore, each of the cultural groups/nations
can be described with the help of “cultural di-
mensions”. In this paper, we use cultural group
and dimension terms interchangeably. To create
three main cultural groups, including MA, LA
and RE, we combine users from several countries.
With human annotation, dimension codes where
assigned to the countries which were analyzed,
as seen in Figure 2.
Figure 2: Countries and Assigned Dimension Codes
3.1 Experimental Setup
Selected Users. Using Twitter Streaming API,
we collected a sample of of 4250109 tweets published
by 3198307 users in the period from 17th to 18th of
March 2014. Only about 2% out of these tweets were
provided with geographic locations. From this “Sam-
ple” dataset, we randomly selected 20.000 users with
published tweets containing geographic locations as
seen in Figure 1. The geographic information avail-
able in the tweets’ meta-data helped to reduce pre-
processing time. In order to accurately determine a
user’s origin, it was paramount to decrease as much
of possible the number of users travelling or resid-
ing in countries other than their country of origin.
For this, we introduced a parameter ↵, which equals
one when user language defined in the user profile
matches with the top first native language related
to the user’s country, and zero otherwise. This re-
quires however a large number of users when training
our classification models for improving accuracy when
identifying users and their respective origin.
User Profiling Features. For the selected users,
we followed their tweets, replies and retweets of their
followers in the period from 18th to 25th of March
2014. The “Follow” dataset comprised these tweets,
in which around 31% of 2853719 tweets are geograph-
ically tagged and originated from the initially selected
users and 131174 of their followers.
For each of the pre-selected geo-tagged users, we
created feature vectors representing a summary of
users’ activity summaries, which were stored into
the dataset “Profiles”. The “Profiles” dataset con-
sists of the 13289 pre-selected users defined as com-
ing from the top most active countries in our sample
dataset, namely the USA (US), Brazil (BR), Indone-
sia (ID), Great Britain (GB), Turkey (TR), Japan
(JP), France (FR), Spain (ES), Malaysia (MY), Mex-
ico (MX), Russia (RU), Canada (CA), Italy (IT) and
where content from followers was available. We ex-
ploited the following features for detecting user coun-
try origin:
•LANGUAGE User Language from user’s Twit-
ter profile.
•BEHAVIOR features set included features re-
lated to user activities on Twitter.
1. Tagging: number of Hashtags divided
by the sum of Hashtags and Uniform re-
source locators (URLs) occurring in user-
generated content; denotes users’ prefer-
ences towards tagging and sharing content
activities.
2. Languages: number of di↵erent languages
detected from user content1, normalized by
division of the mean values of languages em-
ployed by all users.
3. Weekends : number of tweets published
on weekends divided by number of tweets
published by the user; denotes frequency of
posting during weekends.
4. Response: number of user replies divided
by the total number of user replies and
retweets.
5. Mentions: defines user preferences for
sharing information on other users as com-
pared to sharing Hashtags and URLs; cal-
culated from the total number of user men-
tions divided by the sum of URLs and Hash-
tags. This feature reflects users’ focus on
people or organizational activities as de-
scribed in Lewis [29].
6. Mobility: denotes the number of di↵er-
ent country occurrences in the user tweets’
meta-data divided by the mean value of the
number of country occurrences in the tweets
of all users.
7. Timezones: number of di↵erent time
zones in the tweets’ meta-data of a user.
8. Influence: number of Followers divided by
the total number of Followers and Friends.
•META. The text string Meta is created by join-
ing strings of the language code defined in the
user profile, most used Time zone and Location
found in the tweets’ meta-data.
•CONTENT: for each user one tweet’s text con-
tent is extracted.
1Python library “langid”: https://github.com/sa↵sd/langid.py
•LOCATION: location field found in the user
meta-data.
•FOLLOWERS features set consists of features
extracted from the user follower networks.
1. FCountry: the country mentioned most in
followers’ meta-data.
2. FCountries: the number of di↵erent coun-
tries referred to in the followers’ meta-data.
3. FLanguage: the language most mentioned
in the followers’ tweets’ meta-data.
4. FLanguages: number of di↵erent lan-
guages found in the followers’ tweets’ meta-
data.
5. FTimezone: time zone most referred to in
the followers’ meta-data.
6. FTimezones: number of di↵erent time
zones referred to in the followers’ meta-
data.
7. FInfluence: number of Followers divided
by the total number of Followers and
Friends for each of the user Followers, fur-
ther taking the mean value for the user we
follow.
For determining a user country location, we em-
ployed Twitter meta-data and Twitter specific el-
ements. We did not apply named-entity recogni-
tion algorithms, since they are resource-demanding
and some named entities may be quite ambiguous
in distinguishing from other words [30, 31]. The
initial feature choice was guided by the literature
sources and our experimental system design. In ad-
dition to content and user-related features, we also
included follower-related behavior, to examine a fea-
ture’s importance in relation to country detection
performance. In further experiments, we also ex-
amined the user responses in friend-follower relation-
ships.
Country Detecting Classifier. Each of our ini-
tially selected users had an associated country code
(string with ISO 3166-1 alpha-2 country code) found
in the user meta-data of tweets. We used these users
for training our classification model based on the fea-
tures described in this section.
To each of LANGUAGE codes we assigned a nu-
merical value, while the BEHAVIOR feature set in-
cluded only numerical values. In the FOLLOW-
ERS features’ set, we coded FCountry, FLanguage,
FTimezone as numerical values, while the rest
were computed as integers (FCountries, FLanguages,
FTimezones) or real values (Influence). When dealing
with numerical values, for building our classification
models, we exploited the Decision Trees Classifier,
which also allowed to consider the importance of a
feature.
When dealing with language features extracted
from the user tweets (CONTENT, LOCATION and
META), we created pipelines performing the follow-
ing steps:
•Convert text data such as tweets’ content or lo-
cation field to a matrix of token counts;
•Convert the count matrix into its normalized
Term Frequenc y - I nverse Document Frequenc y
statistics;
•Employ the Multinomial Naive Bayes classifier
for predicting user countries.
For evaluating our country classification results, we
performed a three-fold Cross-Validation (CV), and
employed Accuracy and F1-measure [32]. For creat-
ing our training and test samples, we split the “Pro-
files” dataset into fractions of 75% and 25%.
Communication Patterns and User Re-
sponses. Based on the “Follow” dataset, we created
the “Communication” dataset representing 107960 of
user tweets, replies and re-tweets, from which about
were 88% geographically tagged and published by the
pre-selected users and their followers. We then ana-
lyzed user interactions among di↵erent user groups.
Next, we created and evaluated a response predict-
ing classifier based on the ratings computed using the
“Communication” dataset as follows:
•For each of the pre-selected users, we com-
puted the number of their followers’ retweets and
replies, which we exploited for creating a rating
score ranging from 0 (no replies and retweets for
a particular user and follower combination) to 1
(the maximal number of retweets and replies for
user and follower communications). We rated
each pre-selected user in a way to assess the in-
terest for a particular follower.
•We predicted user responses towards their
friends by training our decision tree and logistic
regression models based on the aforementioned
features and a set of features denoting users’
matches in the language usage and location.
4 Results
4.1 Explicit Locations in Meta-data
Figure 3: Countries Detection Test
We were interested in evaluating how useful country-
related information extracted from user tweets’ and
followers’ tweets is. The aim was to find out the re-
quired amount of Twitter-content to infer country-
locations of users in our dataset and assess the possi-
bility of detecting user origins based on the meta-data
of tweets of the user and followers.
Our first hypothesis states that meta-data of user
and followers tweets are not sufficient for explicitly
inferring country locations for a majority of selected
users. The average number of tweets per user is about
119 tweets (Standard deviation 156 tweets), while the
number of friends and followers is about 590 (St.dev.
2731 friends) and 894 (St.dev. 7598 followers) respec-
tively per user in average. Even though the fraction
of geo-tagged tweets from our “Follow” dataset ex-
ceeds the randomly selected tweets from the “Sam-
ple” dataset in more than 15 times, reaching around
31%, the amount of tweets collected during one week
was not sufficient to infer country-specific informa-
tion for about 70% of the users, as shown in Figure
3. In spite that geographically-tagged tweets occur
only in about 2% of cases in our “Sample” dataset,
the location field enabling users to specify their lo-
cations arbitrary in text was filled in 54.6% of cases.
However, here we might disregard the location field
(further examined below in Table 2) and usage of geo-
coding services due to the ambiguity and often use of
the location field by Twitter users for personal or hu-
moristic comments, as stated in [1] in more than 30%
of cases. Therefore, we assert in our H1.1 hypothesis,
that for the majority of cases the country location
meta-data is not sufficient for inferring user countries
based on followers’ and user tweets meta-data alone.
4.2 User Country Prediction
Next, we analyzed how the di↵erent combination of
features can be useful for distinguishing between dif-
Table 2: Performance of the Country-detecting Classifier (in test/train split: 4650 training and 1550 testing
instances, CV Acc.: 3-times cross-validation accuracy for 6200 instances in training and test datasets in total)
CV Acc. Accuracy Precision Recall F1
Feature Set ↵=1 8↵ ↵=1 8↵ ↵=1 8↵ ↵=1 8↵ ↵=1 8↵
User-related Data
LANGUAGE 0.88 0.76 0.88 0.76 0.78 0.70 0.88 0.76 0.82 0.71
LOCATION 0.62 0.58 0.64 0.59 0.63 0.66 0.64 0.59 0.53 0.51
META 0.91 0.85 0.91 0.85 0.90 0.86 0.91 0.85 0.90 0.83
BEHAVIOR+LANGUAGE 0.80 0.66 0.81 0.66 0.81 0.67 0.81 0.66 0.81 0.67
CONTENT 0.63 0.58 0.65 0.58 0.55 0.45 0.65 0.58 0.54 0.47
BEHAVIOR 0.44 0.38 0.46 0.38 0.48 0.39 0.46 0.38 0.47 0.39
Follower-related Data
FOLLOWERS 0.88 0.84 0.88 0.84 0.88 0.84 0.88 0.84 0.88 0.84
Mixed Data
LANGUAGE+FOLLOWERS 0.94 0.87 0.94 0.87 0.94 0.87 0.94 0.87 0.94 0.87
BEHAVIOR+FOLLOWERS 0.87 0.82 0.88 0.83 0.88 0.83 0.88 0.83 0.88 0.83
BEHAVIOR+FOLLOWERS+LANGUAGE 0.92 0.87 0.91 0.87 0.91 0.87 0.91 0.87 0.91 0.87
ferent countries of user origins.
User Data. Our aim was to achieve an accuracy
of above 46% of country prediction when considering
the majority class classifier’s threshold, since about
46% of users in the “Profiles” dataset originated from
the USA, with English defined in their user profiles.
Moreover, in our dataset some of the countries such
as Indonesia (ID) and Malaysia (MY) have a large
fraction of English language users, even though En-
glish is not the native language. This is why some
of the users are misclassified if only the language de-
fined in their user profiles is considered. However, our
findings revealed that the LANGUAGE classification
strategy outperformed all other User-related strate-
gies, for the exception of META, in respect to CV
accuracy, as seen in Table 2.
The META feature is represented by a text string
comprising language defined in user profile, majority
time zone and location. With the META feature, we
achieved 91% CV accuracy (three-fold CV showed the
best CV accuracy performance in all our tests), which
outperformed the performance of the LANGUAGE’s
strategy. Since the location-specific information is
not always accurate [1], we further analyzed other fea-
ture mixes. The CONTENT strategy slightly outper-
formed LOCATION-based strategy in alf a = 1 cases
for Accuracy, Recall and F1 measures while enabling
to achieve 63% CV accuracy when based on only one
tweet’s content per user. The BEHAVIOR strategy
performed poorly compared with other classification
strategies, however, we noticed a slight precision im-
provement in alf a = 1 cases with lower recall and ac-
curacy values when using BEHAVIOR together with
the LANGUAGE feature. It seems that the BEHAV-
IOR feature does not allow us to improve user clas-
sification into country groups when using only user-
related data. Considering a relatively small number
of users in our “Profiles” dataset, we were not sur-
prised to achieve only 44% of CV accuracy and 46%
test accuracy when using the BEHAVIOR feature set,
which might require more data and features to yield
similar results as the strategies analyzed above.
Followers Data. In cases when user meta-
data was not available or deemed to be inaccurate,
the FOLLOWERS strategy could compete with some
User-related feature sets. We achieved more than 5%
improvement in Precision and F1-measure compared
to all User-related feature combinations, with excep-
tion of META. Interestingly, in a majority of test
cases, we observed better performance when consid-
ering ↵= 1 cases. We might reasonably assume, that
the combination of user language and user country
of origin is important for selecting training instances.
Overall, the FOLLOWERS feature set provided a vi-
able alternative for detecting user countries in our
experiments when accurate META and LANGUAGE
data was absent.
User and Followers. When combining user
LANGUAGE with information extracted from the
followers’ network, we achieved the best performance
in all measures, in all our experiments for alf a =
1 cases. The cross-validation accuracy of LAN-
GUAGE+FOLLOWERS combination was around
94%. Therefore, we might accept our hypothesis
H2.1, that the user contacts’ network improves user
country predictions. Despite of our expectations,
combining BEHAVIOR with FOLLOWERS features
did not improve classification performance compared
Table 3: Relative Features Importance (User Country Prediction)
BEHAVIOR+LANGUAGE+FOLLOWERS
Feature Importance
(%)
Feature Importance
(%)
Feature Importance
(%)
Feature Importance
(%)
Language 100 FCountry 16.16 FTimezone 15.36 FInfluence 2.34
Weekends 2.34 Influence 2.13 Mentions 1.55 Tagging 0.98
Response 0.73 FTimezones 0.51 Timezones 0.36 Languages 0.33
FLanguage 0.32 FLanguages 0.26 FCountries 0.15 Mobility 0.02
to using only FOLLOWERS. We could not accept the
H2.2, that the BEHAVIOR patterns could help in im-
proving user country predictions in our experimental
settings. Overall, when using Follower-related fea-
tures and LANGUAGE, we were able to outperform
the accuracy when using only LANGUAGE/META
features and also the Calgari algorithm using Naive
Bayes classification model, as described in [1].
Combining all non-content features in the “BE-
HAVIOR+LANGUAGE+FOLLOWERS” classifica-
tion strategy enabled us to assess the relative impor-
tance of features for country detection using Deci-
sion Tree classifier. Our experimental results showed
that user language, followers’ majority country and
time zone, followers’ average influence and posting
day were the most important features for detecting
user country of origin, while the user mobility was
the least important feature to consider, as seen in
Table 3.
4.3 Predicting Follower Responses
Figure 4: Communication between Cultural Dimen-
sions for geographically-tagged users
Our analysis of communication between users and
their followers demonstrated that a substantial pro-
portion of users follow other users from the same cul-
tural groups as seen in Figure 4. This might indicate,
that our users are more interested to follow others in
their specific area of interest, in particular countries.
Next, we investigated whether the dimension or
country location of user and follower match, as
well as other user-related features having an influ-
ence on communications between Twitter friends.
For this, we created a classification model based
on the decision tree and logistic regression tech-
niques while using the aforementioned 16 features.
We also added binary variables such as “Lang-
Match”,“CountryMatch”,“DimMatch” (True when
user and followers’ profile languages, Countries and
Dimensions match), “FLangMatch” (True when user
followers’ and followers followers’ profile languages
match), “FCMatch” (True when user followers’ and
followers followers’ Country match), ”FTimezMatch”
(True when user followers’ and followers followers’
time zones match).
When a followers’ location was unknown, we em-
ployed the previously constructed classification mod-
els “META” and “LANGUAGE”. The parameters
“CountryMatch” and “DimMatch” were set to True
when friends’ parameter matched with the parameter
of the follower based on one of the three values: value
from the profile information, parameter’s value de-
rived using “META” or “LANGUAGE” country de-
tection models, or employing “LANGUAGE” model
for detecting a dimension value directly.
To find out what is important for predicting user
responses to the friends of a user, we performed the
following steps:
•For each of the initially followed users, we calcu-
lated a sum of follower responses (retweets and
replies) for each of their followers, which was fur-
ther normalized by dividing it with the maxi-
mum sum of responses for a follower. This way
we ranked our initial Twitter users with numbers
from 0 (no response) to 1 (maximum number of
response) to account for the “interestingness” of
each of a user’s followers.
•Since a large fraction of our followers had very
few replies, our ratings included more than 98%
Table 4: Relative Features Importance (RFI) in Followers’ Response Prediction Test using Decision Trees,
Logistic Regression Analysis Results (Predicted Logit of Interest) and Logit Marginal E↵ects,
statistically significant (with p<0.05) are shown in bold font
Parameter D.
Tree
Logistic Regression Results, pseudo R2⇡
0.23, sensitivity ⇡62, specificity ⇡72
Marginal E↵ects
RFI Odds
Ra-
tio
zP>
|z|
Std.
Err.
95%
Conf.Int.
dy/dx Std.
Err.
zP>
|z|
95%
Conf.Int.
Intercept 172.7 2.61 0.01 5.15 1.97 1.29 9.02
CountryMatch* 2.22 0.46 -1.48 0.14 -0.78 0.53 -1.82 0.25 -0.14 0.09 -1.49 0.14 -0.33 0.04
DimMatch* 3.91 4.45 2.31 0.02 1.49 0.64 0.23 2.76 0.23 0.11 2.35 0.02 0.04 0.49
FCMatch* 9.63 0.68 -1.96 0.05 -0.38 0.19 -0.76 -0.00 -0.07 0.03 -1.98 0.05 -0.14 -0.00
FLangMatch* 100 0.09 -10.03 0.00 -2.41 0.24 -2.88 -1.94 -0.43 0.03 -14.23 0.00 -0.49 -0.37
FTimezMatch* 7.20 1.20 0.86 0.39 0.18 0.21 -0.23 0.59 0.03 0.04 0.86 0.39 -0.04 0.11
LangMatch* 6.50 0.79 -0.67 0.50 -0.24 0.35 -0.93 0.46 -0.04 0.06 -0.67 0.50 -0.17 0.08
FCountries 24.79 1.16 1.44 0.15 0.15 0.11 -0.05 0.36 0.03 0.02 1.45 0.15 -0.01 0.06
FInfluence 74.49 0.26 -2.15 0.03 -1.36 0.63 -2.59 -0.12 -0.23 0.11 -2.17 0.03 -0.46 -0.02
FLanguages 56.84 0.85 -3.05 0.00 -0.16 0.05 -0.26 -0.06 -0.03 0.01 -3.12 0.00 -0.05 -0.01
FTimezones 25.78 0.98 -0.76 0.44 -0.02 0.03 -0.07 0.03 -0.00 0.01 -0.77 0.00 -0.01 0.01
Influence 70.61 0.94 -0.08 0.93 -0.06 0.73 -1.49 1.37 -0.01 0.13 -0.08 0.93 -0.27 0.25
Languages 23.37 0.78 -2.88 0.00 -0.24 0.08 -0.41 -0.08 -0.04 0.01 -2.94 0.00 -0.07 -0.01
Mentions 2.08 0.64 -0.62 0.53 -0.45 0.72 -1.86 0.96 -0.08 0.13 -0.62 0.53 -0.34 0.17
Mobility 2.96 0.14 -1.78 0.07 -1.97 1.10 -4.13 0.19 -0.35 0.20 -1.8 0.07 -0.74 0.03
Response 28.60 1.08 0.05 0.96 0.08 1.59 -3.04 3.20 0.01 0.29 0.05 0.96 -0.55 0.58
Tagging 27.93 1.11 0.31 0.75 0.10 0.32 -0.53 0.73 0.2 0.06 0.31 0.75 -0.09 0.13
Timezones 10.82 0.69 -1.65 0.10 -0.37 0.22 -0.81 0.07 -0.07 0.04 -1.66 0.09 -0.14 0.01
Weekends 77.61 2.04 1.37 0.17 0.71 0.52 -0.30 1.73 0.13 0.09 1.38 0.17 -0.05 0.31
with the highest rank of “interestingness”. Out
of 106775 ranks, only 343 ranks were of 0 value
(no interest). This is why for creating our model
based on the 343 ranks of 0 value (no interest),
and 343 ranks of 1 value (highest interest), we
disregarded other ranks.
•Next, having the Rank values as dependent vari-
ables indicating user response to a user’s friend,
user-related and follower-related features, we cre-
ated our classification models.
Decision Trees. Having 686 instances in our
“Interests” dataset, we splited it into training (75%)
and testing (25%) sets for evaluating our classifica-
tion model based on the decision trees technique.
It achieved 65% of accuracy, 69% precision, 63% of
recall and 65% of F1-measure. Even though the
classification accuracy using “leave-one-out” cross-
validation technique, was outperforming a random
classifier in only about 15%, it enabled us to calculate
variables’ importance presented in Table 4 (Features
Importance column) based on the training set of 685
instances.
Surprisingly, features relative importance statis-
tics based on decision trees presented CountryMatch,
DimMatch and LangMatch within the five least im-
portant features set. This is why we could not ac-
cept our Hypothesis H3.1, stating that user and fol-
lowers country locations and languages match are
amongst the most important prediction parameters
for the user responses. The most important features
were FLangMatch, Weekend, FInfluence, Influence,
and FLanguages. It seems that Country and Dimen-
sion match were not as important for the user re-
sponse predictions, when using decision trees classifi-
cation model. We explain this by the possible biases
in our dataset towards the most active users, includ-
ing also broadcasting agencies. To further assess vari-
ables likelihood and e↵ect on users’ interest towards
their friends, we perform logistic regression and com-
pute their marginal e↵ects.
Logistic Regression. We performed logistic re-
gression analysis with the Statsmodel Package2. Ta-
ble 4 presents logistic regression results considering
binary dependent variable of user interest in user’s
friend coded as 0 (no responses) and 1 (most of re-
sponses). Some of the explanatory variables (marked
2http://statsmodels.sourceforge.net/
with *) where categorical and coded as True or False
when the compared values were matched or not. The
overall model was statistically significant with log
likelihood ratio p-value less than 0.001. The pseudo
R here cannot be interpreted as a measure of variance
such as in a least squares regression.
The logistic regression showed the significance of
the FLangMatch and FInfluence included in the top
most important features derived with decision trees.
Interestingly, DimMatch and Languages were also
statistically significant. Marginal e↵ects statistics
presented in the last three columns in Table 4 showed
that DimMatch is associated (statistically significant)
with 23% higher probability of user response. FLang-
Match, FInfluence and FLanguages are related to sta-
tistically significant lower probability of replying in
43%, 23% and 3% respectively. However, Influence
was statistically insignificant in our logistic regres-
sion model. This is why we could not strictly accept
our Hypothesis H3.2.
Hypothesis Revisited and Discussion. One
of the objectives of this work was to investigate the
possibility of exploiting microblogs’ to detect a home
country of a user. For this, we considered sev-
eral countries on Twitter, whose users were deemed
most active in our sample set. Firstly, we observed
that country-name metadata of Twitter users was
matched with the country-name meta-data of their
followers in only 30% of our users. Therefore, we
agreed with our H1.1 hypothesis stating that in most
cases a country location taken from a user and follow-
ers’ meta-data is not sufficient for providing cues on
user origins in our dataset. Secondly, we found out
that the information publicly available in user meta-
data and followers’ network enabled us to predict user
country locations with a considerable accuracy of 90%
or more for the best feature selection strategies we an-
alyzed (RQ1). The most successful feature combina-
tions included both elements, Followers-related data
FOLLOWERS, and User-related data LANGUAGE
(RQ2).
However, usage of BEHAVIOR together with LAN-
GUAGE and FOLLOWERS feature sets did not pro-
vide any improvement. Therefore, we cannot strictly
accept H2.2 without considering other features tak-
ing part in the classification strategy. Nevertheless,
a solution to address the need for user profiling for
improved user experience online, we might suggest
exploiting user behavioral patterns or other well-
performing features set combinations instead of di-
rectly asking user locations. This way, we could sat-
isfy user preferences towards sharing content, times
and ways of communicating with other users, whilst
respecting privacy.
Overall, our results reveal a global orientation of
our Twitter users in our dataset. Country Match
and Language Match were not highly ranked (rel-
ative importance in decision trees), neither statisti-
cally significant features (logistic regression results).
This is why we could not accept hypothesis H3.1.
Dimension Match was more important than Coun-
try Match, and also statistically significant, leading
to improved probability of user replies. Interestingly,
user Influence was ranked in the top of feature impor-
tance in our decision tree results, while logistic regres-
sion showed no statistical significance, and FInfluence
was even more important and significant. Thus, users
with more influential followers might get less replies.
When users have followers, with matching majority
language, their reply probability decreases by 43%,
which still supported the finding on the global nature
of communication on Twitter. However, DimMatch
could not be underestimated and requires a further
investigation in further friend or related content rec-
ommendation experiments.
Privacy. Even though the country and user
cultural dimension detection experiments open pos-
sibilities for adaptation, also concerns in regard of
privacy issues could be raised, in particular for
users with open profiles in Social Networks. Mi-
croblogging meta-data, followers’ network and user-
generated content enabled us to predict user country
locations with considerable accuracy. Avoiding shar-
ing location information in Twitter meta-data might
help in preserving user whereabouts only to a cer-
tain extent. Even language mined from user content
or defined in the user profile provides an insight into
human countries of origin. Therefore, for a privacy-
concerned user, we recommend to withdraw from mi-
croblogging or closing open profiles.
Sampling Biases. It is important to mention that
our data collection method is prone to sampling bi-
ases. Using Twitter sample, we might be biased to-
wards the most active users such as event or news
broadcasters, which requires further analysis.
On Sociological Models Usage. The Internet
brings users from diverse cultures together, however,
understanding their needs and requirements in order
to realize quality features for web applications is chal-
lenging. Applying sociological models to assess web
site quality as perceived by users might not be trivial
due to globalization, since the new e-culture of indi-
viduals often does not comply with rules described in
models referring to di↵erences in cultural personality
[33]. Therefore, we might require new approaches to
study user behavior online while respecting privacy.
Limitations. Our sample set contains users from
the most active countries, providing geographical lo-
cations in meta-data. The user-generated content was
collected for a period of one week. It is reasonable
to assume that real-life activities might a↵ect user
behaviors. This is why we plan to explore user mi-
croblogging activities while assessing di↵erent models
and feature combinations for predicting user origins
and communication patterns in order to evaluate our
approach in a larger time frame and extended loca-
tions set. For a thorough evaluation of our approach
in predicting user responses to their friends, we might
also involve human assessors working with our online
version of the prototype.
5 Conclusions
In this paper, we analyzed microblogging activities
for persons from the top 13 most active countries on
Twitter. We investigated di↵erent feature sets ex-
tracted from the microblogging content and meta-
data of publicly available tweets. Our findings re-
veal that combining user-related microblogging fea-
tures and features extracted from a followers’ net-
work enables user country prediction with an accu-
racy of more than 90%. We reflected on the results,
also in view of human privacy issues, and provided
recommendations for users concerned about sharing
their data in open microblogs. Considering sociolog-
ical studies and previous works on behavioral di↵er-
ences online, we proposed an approach for mining
individual culture-specific microblogging preferences
which we abstracted from country information, often
revealed in user meta-data and content. This pro-
vides insights on the adaptation for web applications’
and personalization options in order to preserve hu-
man privacy, while improving user satisfaction online
by providing application features/content which are
of interest for the cultural origins of the user. Fi-
nally, we investigated user interactions, and found out
that users from the same cultural groups tend to com-
municate more with each other. In our next paper,
we will analyze user communication amongst cultural
groups in the long term, aiming to uncover recom-
mendation approaches for further improving user ex-
perience on the Web. The other Twitter-specific fea-
tures we might explore include tweeting frequency,
topics found in the tweets’ content, and an in depth
tagging behavior analysis.
References
[1] B. Hecht, L. Hong, B. Suh, and E.H. Chi, “Tweets
from Justin Bieber’s Heart: the Dynamics of the
Location Field in User Profiles”, in Proc. of the
SIGCHI Conference on Human Factors in Com-
puting Systems. ACM, 2011, 237–246.
[2] B. Han, P. Cook, and T. Baldwin, “Text-based
Twitter User Geolocation Prediction.”, J. Artif.
Intell. Res.(JAIR), 2014, vol. 49, 451–500.
[3] L. Ghahremanlou, W. Sherchan, and J.A. Thom,
“Geotagging Twitter Messages in Crisis Manage-
ment”, The Computer Journal, 2014, bxu034.
[4] Z. Cheng, J. Caverlee, and K. Lee, “You are
Where you Tweet: a Content-based Approach to
Geo-locating Twitter Users”, in Proceedings of
the 19th ACM International Conference on Infor-
mation and Knowledge Management. ACM, 2010,
759–768.
[5] J. Hare, J. Davies, S. Samangooei, and P. Lewis,
“Placing Photos with a Multimodal Probability
Density Function”, in Proceedings of Interna-
tional Conference on Multimedia Retrieval.ACM,
2014, 329.
[6] D. Jurgens, “That’s What Friends are for: Infer-
ring Location in Online Social Media Platforms
Based on Social Relationships.”, in ICWSM,
2013.
[7] E. Ilina, “A User Modeling Oriented Analysis
of Cultural Backgrounds in Microblogging”, HU-
MAN JOURNAL, 2012, 1 (4), 166–181.
[8] J. Mahmud, J. Nichols, and C. Drews, “Home
Location Identification of Twitter Users”, ArXiv
Preprint ArXiv:1403.2345, 2014.
[9] V. Rakesh, C.K. Reddy, D.Singh, and M.S. Ra-
machandran, “Location-specific Tweet Detec-
tion and Topic Summarization in Twitter”, in
Advances in Social Networks Analysis and Min-
ing (ASONAM), 2013 IEEE/ACM International
Conference on. IEEE, 2013, 1441–1444.
[10] Y. Chen, J. Zhao, X.Hu, X. Zhang, Z. Li, and
T.-S. Chua, “From Interest to Function: Location
Estimation in Social Media.”, in AAAI, 2013.
[11] J. Gelernter and W. Zhang, “A Learning Method
to Geocode Location Expressions in Twitter Mes-
sages”, Journal of Spatial Information Science,
2014.
[12] S. Hale, D. Ga↵ney, and M. Graham, “Where in
the World are You? Geolocation and Language
Identification in Twitter”, Professional Geogra-
pher.
[13] J. Lingad, S. Karimi, and J. Yin, “Location
Extraction from Disaster-related Microblogs”, in
Proceedings of the 22nd International cCnference
on World Wide Web companion. International
World Wide Web Conferences Steering Commit-
tee, 2013, 1017–1020.
[14] S. Roller, M. Speriosu, S. Rallapalli, B. Wing,
and J. Baldridge, “Supervised Text-based Geolo-
cation Using Language Models on an Adaptive
Grid”, in Proceedings of the 2012 Joint Confer-
ence on Empirical Methods in Natural Language
Processing and Computational Natural Language
Learning. Association for Computational Linguis-
tics, 2012, 1500–1510.
[15] O. Van Laere, J. Quinn, S. Schockaert, and B.
Dhoedt, “Spatially Aware Term Selection for
Geotagging”, Knowledge and Data Engineering,
IEEE Transactions on, 2014, vol. 26, 1, 221–234.
[16] N. Ireson and F. Ciravegna, “Toponym Reso-
lution in Social Media”, in The Semantic Web–
ISWC 2010, Springer, 2010, 370–385.
[17] R. Compton, M.S. Keegan, and J. Xu, “Inferring
the Geographic Focus of Online Documents from
Social Media Sharing Patterns”, ArXiv Preprint
ArXiv:1406.2392, 2014.
[18] Y. Takhteyev, A. Gruzd, and B. Wellman, “Ge-
ography of Twitter Networks”, Social networks,
2012, vol. 34 (1), 73–81.
[19] J. Kulshrestha, F. Kooti, A. Nikravesh, and P.K.
Gummadi, “Geographic Dissection of the Twitter
Network.”, in ICWSM, 2012.
[20] B.A. Olaniran, “Culture, Learning Styles, and
Web 2.0”, Interactive Learning Environments,
2009, vol. 17 (4), 261–271.
[21] B.E. Wiggins, The Impact of Cultural Dimen-
sions and the Coherence Principle of Multimedia
Instruction on the Achievement of Educational
Objectives within an Online Learning Environ-
ment, PhD thesis, Indiana University of Penn-
sylvania, 2011.
[22] T.H. Silva, P. de Melo, J. Almeida, M. Mu-
solesi, and A. Loureiro, “You are What you Eat
(and Drink): Identifying Cultural Boundaries by
Analyzing Food & Drink Habits in Foursquare”,
ArXiv Preprint ArXiv:1404.1009, 2014.
[23] R. Garcia-Gavilanes, D. Quercia, and A. Jaimes,
“Cultural Dimensions in Twitter: Time, Individ-
ualism and Power”, Proceedings of the 9 th AAAI
ICWSM, 2013.
[24] R. Garcia Gavilanes, Neil OHare, Luca Maria
Aiello, and Alejandro Jaimes, “Follow my Friends
this Friday! An Analysis of Human-generated
Friendship Recommendations”, in Social Infor-
matics, 2013, Springer, 370–385.
[25] M. Giatsoglou, D. Chatzakou, and A. Vakali,
“Community Detection in Social Media by Lever-
aging Interactions and Intensities”, in Web
Information Systems Engineering–WISE 2013,
Springler, 2013, 57–72.
[26] Y. Artzi, P. Pantel, and M. Gamon, “Predicting
Responses to Microblog Posts”, in Proceedings of
the 2012 Conference of the North American Chap-
ter of the Association for Computational Linguis-
tics: Human Language Technologies. Association
for Computational Linguistics, 2012, 602–606.
[27] J. Schantl, R. Kaiser, C. Wagner, and M.
Strohmaier, “The Utility of Social and Topical
Factors in Anticipating Repliers in Twitter Con-
versations”, in Proceedings of the 5th Annual
ACM Web Science Conference. ACM, 2013, 376–
385.
[28] M.D.Myers and F.B. Tan, “Beyond Models
of National Culture in Information Systems Re-
search”, Advanced Topics in Global Information
Management, 2003, vol. 2, 14–29.
[29] R.D. Lewis, When Cultures Collide: Manag-
ing Successfully Across Cultures, Nicholas Brealey
Publishing, 2000.
[30] J. Nothman, N. Ringland, W. Radford, T. Mur-
phy, and J.R. Curran, “Learning Multilingual
Named Entity Recognition from Wikipedia”, Ar-
tificial Intelligence, 2013, vol. 194, 151–175.
[31] D. Nadeau and S. Sekine, “A Survey of Named
Entity Recognition and Classification”, Lingvisti-
cae Investigationes, 2007, vol. 30 (1), 3–26.
[32] T. Fawcett, “Roc Graphs: Notes and Practical
Considerations for Researchers”, Machine learn-
ing, 2004, vol. 31, 1–38.
[33] M. Sigala and O. Sakellaridis, “Web Users’ Cul-
tural Profiles and E-service Quality: Internation-
alization Implications for Tourism Web Sites”, In-
formation Technology & Tourism, 2004, vol. 7 (1),
13–22.
