This document discusses a research project aimed at understanding how cultural and geolocation factors influence user behavior on Twitter. The study explores the potential to leverage these insights for better friend and content recommendations through predictive models based on user profiles and communication patterns. Key findings emphasize the importance of language, user influence, and geographical data in predicting follower responses and country origins.

1. Cultural and Geolocation
Aspects of Communication
in Twitter
Elena&Daehnhardt1&
Yanguo&Jing2&
Nick&Taylor1&
&
1.&Heriot;Wa=&University&&&&&&&&&&&&2.&London&Metropolitan&University&

2. About the Project
The$main$goals:$to$find$out$how$cultural$differences$
impact$user$behaviour$on$Twi9er$and$exploit$these$
differences$in$adapta;on:$
&
• Cultural&Behaviour&Pa=erns&in&Microblogs&
• Communica;on$in$Twi9er$(this$paper)$
• ExploiIng&Cultural&Differences&in&Recommenders&&
(further&work)&
2

3. Outline
• Research&QuesIons&
• Approach&
• Experimental&Setup&
• Findings&
• Further&Work&
3

4. The Main Idea
4
• To$understand$if$communica;on$on$Twi9er$is$
influenced$by$user$origins.$
• Why?&PotenIally,&we&could&use&this&knowledge&
for&friends/content&recommendaIons&(TODO).&
• What&we&need?&Country&locaIons&are&required.&
• In&the&previous&paper,&cultural&microblogging&
pa=erns&enabled&to&predict&user&countries & &&
;&challenge&for&a&larger&countries&set.&

5. Research Questions
5
• How&we&can&exploit&Twi=er&to&infer&user&
countries?&
• Which&features&are&important&for&inferring&
countries?&
• Which&friend&features&are&important&for&
predicIng&follower&responses?&

6. “Follow”$
(tweets)$
PredicIon&
Model&
Follow&Selected&
Twi=er&Users&
Store&Tweets&of&
Users&and&Their&
Followers&
Create&User&
Profiles&of&the&
Selected&Users&
User$
Profiles$
Create&Country&
PredicIon&
Model&
Predict&Follower&
LocaIons&and&Create&
their&Profiles&
FOLLOWERS$
Approach: Data Collection & Country Prediction

7. “Follow”$
(tweets)$
User$
Profiles$
Create&
CommunicaIon&
Dataset&
“CommunicaIon”&
Users&&&Follower&Responses&
FOLLOWERS$
Approach: Response Prediction & Evaluation
Significance&of&User&
and&Follower&
Features&
Create&and&Evaluate&
PredicIon&Models&of&
Follower&Responses&

8. Geographic Locations
8
LinearNAc;ve$ US,&UK,&Canada&
ReacIve& Indonesia,&Japan,&Malaysia&
MulI;AcIve& Brazil,&Italy,&France,&Russia,&Spain,&
Turkey,&Mexico&

9. User Profiles
Feature$Sets$ Descrip;on$
LANGUAGE&& User&language&in&the&Twi=er&profile&
BEHAVIOR& Time&of&posIng,&user&influence,&hashtags&usage,&
…&
META& LANGUAGE&+&most&used&Time&zone&and&
LocaIon&
LOCATION& LocaIon&field&found&in&the&user&meta;data.&&
CONTENT& Of&a&one&tweet&
FOLLOWERS$ Most;frequent&country,&Ime&zone&and&average&
followers’&influence,&…&
11
UserNrelated$

10. 12
H1.1.&The&informaIon&on&country&locaIon&
derived&from&user&tweets’&meta;data&and&
respecIve&meta;data&of&the&followers&is&not&
sufficient&for&predicIng&countries&✔&
&

11. Languages in Twitter Profiles
13
Number&of&Users&
%&

12. H2.1& A&users&contact&network&can&assist&in&
improving&country&predicIon.&✔&
H2.2& User&BEHAVIOR&pa=erns&can&assist&in&further&&
improving&locality&predicIon.&✗&
Feature$Set$ CV$Acc.$ Accuracy$ Precision$ Recall$ F1$
User;related&Data&
α=1&& α&& α=1&& α&& α=1&& α&& α=1&& α&& α=1&& α&&
LANGUAGE&
(LANG.)&
.88& .76& .88& .76& .78& .70& .88& .76& .82& .71&
LOCATION& .62& .58& .64& .59& .63& .66& .64& .59& .53& .51&
META& .91& .85& .91& .85& .90& .86& .91& .85& .90& .83&
BEHAVIOR+LANG.$ .80$ .66$ .81$ .66$ .81$ .67$ .81$ .66$ .81$ .67$
CONTENT& .63& .58& .65& .58& .55& .45& .65& .58& .54& .47&
BEHAVIOR&(BEH)& .44& .38& .46& .38& .48& .39& .46& .38& .47& .39&
FOLLOWERS&(FOL)& .88& .84& .88& .84& .88& .84& .88& .84& .88& .84&
Mixed&
Data&
LANG.+FOL.$ .94$ .87$ .94$ .87$ .94$ .87$ .94$ .87$ .94$ .87$
BEH.+FOL.$ .87$ .82$ .88$ .83$ .88$ .83$ .88$ .83$ .88$ .83$
BEH.+FOL.+LANG.$ .92$ .87$ .91$ .87$ .91$ .87$ .91$ .87$ .91$ .87$
Country
Prediction

13. Features Importance
Feature$ Importance$ Feature$ Importance$
Language$ 100$ FCountry$ 16.16$
FTimezone$ 15.36$ FInfluence$ 2.34$
Weekends$ 2.34$ Influence$ 2.13$
MenIons& 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&
16
(Country Predictions)

14. Countries of Repliers
17

15. Predicting Follower Replies
18
Which$friend$features$are$important$for$predic;ng$a$
user’s$follower$responses?$$
$
User&and&follower’s&country&locaIons’&and&language&
match&are&amongst&the&most&important&predicIon&
parameters&for&user&responses.&✗&
&
User’s&influence&is&significant&in&predicIng&her&follower&
responses.&✗&
&

16. Predicting Follower Replies
19
Not Significant Significant
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
z P >
|z|
Std.
Err.
95%
Conf.Int.
dy/dx Std.
Err.
z P >
|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
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-
//&

17. 20
Not Significant Significant
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
RFI Odds
Ra-
tio
z P >
|z|
Std.
Err.
95%
Conf.Int.
dy/dx S
E
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
DimMatch* 3.91 4.45 2.31 0.02 1.49 0.64 0.23 2.76 0.23 0
FCMatch* 9.63 0.68 -1.96 0.05 -0.38 0.19 -0.76 -0.00 -0.07 0
FLangMatch* 100 0.09 -10.03 0.00 -2.41 0.24 -2.88 -1.94 -0.43 0
FTimezMatch* 7.20 1.20 0.86 0.39 0.18 0.21 -0.23 0.59 0.03 0
LangMatch* 6.50 0.79 -0.67 0.50 -0.24 0.35 -0.93 0.46 -0.04 0
FCountries 24.79 1.16 1.44 0.15 0.15 0.11 -0.05 0.36 0.03 0
FInfluence 74.49 0.26 -2.15 0.03 -1.36 0.63 -2.59 -0.12 -0.23 0
FLanguages 56.84 0.85 -3.05 0.00 -0.16 0.05 -0.26 -0.06 -0.03 0
FTimezones 25.78 0.98 -0.76 0.44 -0.02 0.03 -0.07 0.03 -0.00 0
Influence 70.61 0.94 -0.08 0.93 -0.06 0.73 -1.49 1.37 -0.01 0
Languages 23.37 0.78 -2.88 0.00 -0.24 0.08 -0.41 -0.08 -0.04 0
Mentions 2.08 0.64 -0.62 0.53 -0.45 0.72 -1.86 0.96 -0.08 0
Mobility 2.96 0.14 -1.78 0.07 -1.97 1.10 -4.13 0.19 -0.35 0
Response 28.60 1.08 0.05 0.96 0.08 1.59 -3.04 3.20 0.01 0
//&

18. 21
Not Significant Significant
p < 0.05) are shown in bold font
stic Regression Results, pseudo R2
⇡
, sensitivity ⇡ 62, specificity ⇡ 72
Marginal E↵ects
s z P >
|z|
Std.
Err.
95%
Conf.Int.
dy/dx Std.
Err.
z P >
|z|
95%
Conf.Int.
7 2.61 0.01 5.15 1.97 1.29 9.02
-1.48 0.14 -0.78 0.53 -1.82 0.25 -0.14 0.09 -1.49 0.14 -0.33 0.04
2.31 0.02 1.49 0.64 0.23 2.76 0.23 0.11 2.35 0.02 0.04 0.49
-1.96 0.05 -0.38 0.19 -0.76 -0.00 -0.07 0.03 -1.98 0.05 -0.14 -0.00
-10.03 0.00 -2.41 0.24 -2.88 -1.94 -0.43 0.03 -14.23 0.00 -0.49 -0.37
0.86 0.39 0.18 0.21 -0.23 0.59 0.03 0.04 0.86 0.39 -0.04 0.11
-0.67 0.50 -0.24 0.35 -0.93 0.46 -0.04 0.06 -0.67 0.50 -0.17 0.08
1.44 0.15 0.15 0.11 -0.05 0.36 0.03 0.02 1.45 0.15 -0.01 0.06
-2.15 0.03 -1.36 0.63 -2.59 -0.12 -0.23 0.11 -2.17 0.03 -0.46 -0.02
-3.05 0.00 -0.16 0.05 -0.26 -0.06 -0.03 0.01 -3.12 0.00 -0.05 -0.01
-0.76 0.44 -0.02 0.03 -0.07 0.03 -0.00 0.01 -0.77 0.00 -0.01 0.01
-0.08 0.93 -0.06 0.73 -1.49 1.37 -0.01 0.13 -0.08 0.93 -0.27 0.25
-2.88 0.00 -0.24 0.08 -0.41 -0.08 -0.04 0.01 -2.94 0.00 -0.07 -0.01
-0.62 0.53 -0.45 0.72 -1.86 0.96 -0.08 0.13 -0.62 0.53 -0.34 0.17
-1.78 0.07 -1.97 1.10 -4.13 0.19 -0.35 0.20 -1.8 0.07 -0.74 0.03
0.05 0.96 0.08 1.59 -3.04 3.20 0.01 0.29 0.05 0.96 -0.55 0.58
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
RFI Odds
Ra-
tio
z P >
|z|
Std.
Err.
95%
Conf.Int.
dy/dx Std
Err
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 0.0
DimMatch* 3.91 4.45 2.31 0.02 1.49 0.64 0.23 2.76 0.23 0.1
FCMatch* 9.63 0.68 -1.96 0.05 -0.38 0.19 -0.76 -0.00 -0.07 0.0
FLangMatch* 100 0.09 -10.03 0.00 -2.41 0.24 -2.88 -1.94 -0.43 0.0
FTimezMatch* 7.20 1.20 0.86 0.39 0.18 0.21 -0.23 0.59 0.03 0.0
LangMatch* 6.50 0.79 -0.67 0.50 -0.24 0.35 -0.93 0.46 -0.04 0.0
FCountries 24.79 1.16 1.44 0.15 0.15 0.11 -0.05 0.36 0.03 0.0
FInfluence 74.49 0.26 -2.15 0.03 -1.36 0.63 -2.59 -0.12 -0.23 0.1
FLanguages 56.84 0.85 -3.05 0.00 -0.16 0.05 -0.26 -0.06 -0.03 0.0
FTimezones 25.78 0.98 -0.76 0.44 -0.02 0.03 -0.07 0.03 -0.00 0.0
Influence 70.61 0.94 -0.08 0.93 -0.06 0.73 -1.49 1.37 -0.01 0.1
Languages 23.37 0.78 -2.88 0.00 -0.24 0.08 -0.41 -0.08 -0.04 0.0
Mentions 2.08 0.64 -0.62 0.53 -0.45 0.72 -1.86 0.96 -0.08 0.1
Mobility 2.96 0.14 -1.78 0.07 -1.97 1.10 -4.13 0.19 -0.35 0.2
Response 28.60 1.08 0.05 0.96 0.08 1.59 -3.04 3.20 0.01 0.2
//&

19. Findings
User$Country$Predic;on:$
$
• InformaIon&publicly&available&in&user&meta;data&and&
followers’&network&enabled&us&to&predict&user&country&
locaIons&with&an&accuracy&of&>&90%&
• The&most&successful&feature&combinaIons&included&
Followers;related&data&such&as&their&Country,&Timezone&
and&Influence,&and&User;related&data&such&as&Language&
defined&in&User&Profile,&Influence&and&TweeIng&on&
Weekends.&&
& 23

20. Findings
Predic;ng$User$Replies$
$
• Dimension&Match&was&more&important&than&Country&
Match,&and&also&staIsIcally&significant,&leading&to&
improved&probability&of&user&replies.&
• Users&with&more&influenIal&followers&might&get&less&
replies.&&
• When&users&have&followers,&with&matching&majority&
language,&their&reply&probability&has&dropped&by&43%.&
&
24

21. Further Work
• Analyze&microblogging&acIviIes&in&long;term.&
• Evaluate&recommendaIon&strategies&considering&
cultural&origins&of&users.&
• InvesIgate&in&depth&more&content;related&
features&including&contents&of&micro&posts&and&
hashtags&
25

22. Related Paper
E.&Ilina&(Daehnhardt).&A&User&Modeling&Oriented&
Analysis&of&Cultural&Backgrounds&in&Microblogging.&
HUMAN&JOURNAL,&1(4):166–181,&2012.&
Your Questions?
&
&
elena@daehnhardt.com&
& 26