Is the culture of a country associated with the way people use Twitter? The answer is a definite “Yes!” We analyzed more than 2.34 million geo-located user profiles in 30 countries plus their tweets for 10 weeks
Blogpost: http://crowdresearch.org/blog/?p=6767
Paper:http://www.aaai.org/ocs/index.php/ICWSM/ICWSM13/paper/view/6102
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack
Twitter: Time, Individualism and Power
1. Cultural Dimensions in Twitter:
Time, Individualism and Power
Ruth García-Gavilanes (UPF - Barcelona)
Daniele Quercia (Yahoo! Barcelona)
Alejandro Jaimes (Yahoo! Barcelona)
1
8. HOW TO MEASURE CULTURE
• Geert Hofstede: Cultural dimensions
Individualism
Power Distance
• Levine : Pace of Life (Geography of time)
• Perception of time
• Edward T. Hall
Monochronic vs Polychronic
8
10. Culture and Social Media
Can such differences also be captured from
online interactions?
10
11. How to measure culture online?
• Pace of Life
Predictability (tweets, mentions)
Tweets in working hours
• Individualism vs. Collectivism
Users interacting with others (mentions)
• Power Distance : Popularity
Follow, recommend and accept recommendation
preferentially from more popular users
(in-degree imbalance).
11
12. Sampling for 10 weeks in 2011
12
2.34
geolocated
users
follows
#FollowFridday
12.6 K
1.9M
362 K
100>= out-degree <=1000
13. Top 30 countries to study
The top 30 countries by # of users is representative of internet users
13
United States
Brazil
United Kingdom
Indonesia
Canada
Mexico JapanNetherlands
Venezuela SpainAustralia
Germany
FranceArgentina
Chile
South Korea
Colombia
IrelandSouth Africa India
Philippines TurkeyItaly
RussiaSweden
New ZealandNorway
MalaysiaBelgiumSingapore
4
5
6.5 7.0 7.5 8.0 8.5
Log(Internet Penetration)
Log(UserCount)
r = 0.63***
15. Pace of Life : Predictability
Entropy
a) tweets b) mentions
− pi ( j)log2 pi ( j)
j=1
Ni
∑pi ( j):
15
12AM
6AM
9AM
6 PM
9PM
12PM
# tweets in
working hours
Interval j
16. Pace of Life : Predictability
Tweets Mentio
ns
Users in working
hours
Pace of life **-0.62 **-0.68 **-0.58
16
The higher the pace of life , the more predictability
The higher the pace of life the less fraction of users
will tweet during working hours
p < 0:005 (***), p < 0:05 (**), and p < 0.1 (*)
17. Individualism : Interacting with others
Indonesia
Venezuela
Mexico
JapanBrazilColombia
Chile
South Korea Argentina
PhilippinesMalaysia Spain NetherlandsTurkey
South Africa
Singapore Ireland Canada
FranceBelgiumSweden
Norway
New Zealand
Italy
Russia India
Germany
80
85
90
95
100
20 40 60 80
Individualism Index
FractionofEngagement
17
r= ***-0.55
p < 0:005 (***), p < 0:05 (**), and p < 0.1 (*)
Collectivist countries interact more with others
18. Individualism : Interacting with others
Indonesia
Venezuela
Mexico
JapanBrazilColombia
Chile
South Korea Argentina
PhilippinesMalaysia Spain NetherlandsTurkey
South Africa
Singapore Ireland Canada
FranceBelgiumSweden
Norway
New Zealand
Italy
Russia India
Germany
80
85
90
95
100
20 40 60 80
Individualism Index
FractionofEngagement
18
Hong et al.. “Language
matters in twitter:
A large scale study”
ICWSM 11
21. Power Distance: popularity imbalance
Followers Followers/
Followees
Users and followees **-0.62 **-0.67
Users and recommended user **-0.56 **-0.46
User and accepted recommended
user
-0.44 -0.29
21
p < 0:005 (***), p < 0:05 (**), and p < 0.1 (*)
Users prefer to follow and recommend more popular users than
themselves in countries with a higher power distance
24. Why is this important?
Indicator Pace of Time :
Predictibility
Individuali
sm:
Mentions
Power
Distance:
ImbalanceMentions Users (%)
GDP per
capita
***0.55 **-0.57 **-0.41 **-0.48
Education ***0.58 **-0.51 -0.24 ***-0.60
Inequality ***-0.53 **0.49 *0.39 ***0.58
24
25. What is next?
25
• Language dependent features
• More Cultural Dimensions
• Temporal comparisons
More features
26. What is next?
26
• User recommender
• Individualistic vs. collectivistic ?
• Predictable vs. unpredictable ?
• Interfaces personalization
• Do collectivist countries need additional features to
interact easier?
• More engagement?
• Information Propagation
• By knowing the cultural characteristics of users, can
we increase re-tweet chance?
Application
Good afternoon, thanks for coming to my talk. I am an intern at Yahoo Research and student of Universidad Pompeu de Fabra in Barcelona. Today I will talk about “Time, Individualism and Power” which is related to ..
“Cultural Studies” and
and Microblogs (as Janeth), but before entering into the details of this project, let’s give a closer look to the meaning of culture
the concept of culture is used in many ways and it has different meanings, Some use it to refer to civilizations and others as the refinement of the mind relating it to education or art. However, the most widely accepted and used definition of culture in science comes from Geert Hofstedesstudies.
In simple words, culture is a dimension that distinguishes members of one group or categories of people from others. For example, the comparison of behavior between individuals born and raised in the Unites States with individuals born and raised in Japan.
Studying Culture involves studying and comparing the collective behavior learned in different societies as opposed to personality which focused on the study of individuals and their inherited and learned characteristics and as opposedto Human Nature which focused on the Universal inherited characteristics of humans.
Culture influences the choices that we make, how we work, how we interact with others, etc. It is important therefore to take into accountthe cultural differences of groups for several reasons such as in business and marketing. So with this concenpt in mind, can we quantify culture?
The answer is Yes we can! Over several years, a lot has been done to measure culture, specially from these threeanthropologists: Hofstede, Levine and Hall.
Imagine the world divided by cultural dimensions where countries are categorized by the value of each dimension. Pace of LifeThe dimension of pace of life was studied by Levine. It measures how fast and punctual people perform tasks. For example, Switzerland is considered very punctual and fast as opposed to Brazil considered as more spontaneous and relaxed with time. Pace of life is also related to predictability of actions (which reflect the studies of Hall). More punctual and fast countries are considered more predictable than more spontaneous societies due to the tendency of having inflexible schedules and deadlines. ( I said in “general” because there are some exceptions such as Japan that tends to be punctual and fast in carrying out tasks but unpredictable in their actionsdue to their flexibility in changing their schedule (Hall). ) Power DistanceThe dimension of Power Distance measures the extend to which people feel confortable with unequal distribution of power. For example, China and Russia areconsidered countries with high power distance as opposed to USA and Australia.IndividualismThe dimension of Individualism measures the extend to which people are individualistic and maintain weak ties with others. For example, USA and Canada considered as highly individualistic countries as opposed to Indonesia or Brazil.
Knowing that culture in the world has been measured through surveys and experimentsCan we detect differences in pace of life, individualism and power distance in the online world?
We argue that yes, it is possible and to do so we have correlated different twitter metrics to each of the three dimensions (pace of time, individualism and power distance). For Pace of time, we measure whether you are predictable when you tweet and mention others and how many tweets you post as compared to other cultures. For Individualism we measure whether you interact more with others than people from other cultures and for Power Distance we measure whether you preferentially follow, recommend and accept recommendations from other users more popular than you, To understand what I mean with recommendation, lets give a look at the dataset
In order to test our hypotheses, We started by collecting 12.6k geolocated seed users and their followee network during 10 weeks as well as all the tweetsposted from all of these users. From the tweets, we looked at those with the followfridayhashtag, Which is a tag used in 2011 to recommend users to follow. We collected the profile information of the users recommended and We checked at those that were accepted by the seed users. In total we arrive to 2,34 M users studied. By comparing the snapshots of the network of the seeds, we know which recommended user was accepted.
In order to capture enough users, we selected countries with more than 5 K users which accounts for the top 30 countries. We tested the representativeness of the users of these 30 countriesWith the internet penetration per country and find that they are highly correlated. All this information lead us to haveThe time users post tweets adjusted by the corresponding time zoneWho users prefer to followWho recommends who andWho accepts whom through #FollowFridayhashtags
First we have divided every working day in 5 time intervals and for pace of life, we have used the Entropy which measure the unpredictability of a user i posting in interval j (both for tweets and mentions)and the fraction of users posting during working hours (check later if it happens for all time intervals). For the entropy for tweets and mentions, we take the average entropy of all users for each country andcorrelate it with Pace of Life Index. For the # of tweets in working hours, we take the average per user in each country and correlated it with Pace of Life Index.
We find that in all cases, the perason correlations (in this case negative) are significant and also high. It seems then that countries with a higher pace of life tend to be more predictable when they tweet and the higher the pace of life the less fraction of users will tweet during working hours
We move to measure individualism now. Here we take the overall number of usersmentioning others and find a high correlation with their corresponding individualism index.
It is interesting to notice the case of Germany. It has been found that German tweets received the least number ofmentions out of the 10 most common languages in Twitter (Hong, Convertino, and Chi 2011) and that in Germany,few comments are left in blogs, (Mandl 2009).---
Here we present an example of the fraction of users mentioning others for two highly collectivist countries (Brazil and Indonesia) and three highly individualistic countries (USA, Canada and UK)For 5 time intervals in a day. We see that for all intervals, collectivist countries will tend to interact more with others than in individualistic ones.
Now lets move to Power Distance. In our dataset, we have seed users, their followees and users recommended through FF hashtags. Some of whom are befriended by the seeds and some are not. Using these users, we measure power distance through popularity imbalance.We define popularity imbalance as the difference between your popularity and the popularity of who you follow, recommend and accept recommendations. Popularity here has been measured in two ways: by the in-degree value and the fraction between in-degree and out-degree.
We find one more time a high correlation between countries with high power distance and the popularity imbalance ofwho they follow and recommend. Nevertheless, the correlation is not as high and significant for the acceptance ofrecommended users. It seems then that there is a higher popularity imbalance when users follow and recommend other users in countries with higher power distance(in case of questions) Nevertheless, the correlation is not as high for the recommendation acceptance.We explain this due to highly noisy nature of Follow Friday Recommendations.
We also see here that for the majority of the cases, the countries follow the line of the correlation. Cite the study about Indonesian people and they following artists , pop singers, etc.
We also see here that for the majority of the cases, the countries follow the line of the correlation. Cite the study about Indonesian people and they following artists , pop singers, etc.
Why is this important?We go further, by correlating our findings with economic and social indicators which are: GDP per capita (GDP divided by the number of population), Education (the proportion of the GDP invested in Education) and Inequality (as the unequal distribution of family income ). We see that for almost all cases, the findings are are also correlated withGDP per capita, Education and Inequality. An exception is found between education and the interaction with other users.
So what is next?So the most important question now to ask is when and how to apply this knowledge in building applications as to improve the user experience. We are currently studying the use of cultural features in recommender systems of social connections not onlyto increase the probability of acceptance but also to increase the tenure (longevity) of the recommendation in the network of the user. Generally applications now offer equal services without taking into consideration cultural differences. Interfaces could be personalized by considering for example the tendency of users to use twitter to chat and interact with others more than other countries. Cultural differences can also be studied in information propagation, for example does the likelihood of people interacting others increases the change of information propagation?