An´alisis de sen imien o en e en os con
con incan es
Ashley J Na eso C an o d
Di ec o : Ra ael Caballe o Rold´an
2
Resumen
En es e abajo conside amos el p oblema de la ca ac e izaci´on de mensajes y
usua ios en las edes sociales, en es e caso el se icio de mic oblogging Twi e .
Pa a ello hemos desca gado 13,3 millones de wee s u lizando como con ex o las
elecciones gene ales es adounidenses que u ie on luga en el mes de no iemb e
de 2016. A pa i de los wee s desca gados cons uimos un conjun o cohe en e
median e un p oceso de limpieza de da os. Una ez cons uido, analizamos el
conjun o u ilizando la ´ecnica del an´alisis de sen imien o pa a asigna de o ma
au om´a ica una e ique a desc ip i a a cada wee que indica si el au o del wee
apoya a alguno de los candida os o, po lo con a io, se opone ac i amen e a
alguno de los candida os.
Una ez compilados los da os ob enemos una se ie de esul ados, especial-
men e donde analizamos el compo amien o de los usua ios que apoyan a un can-
dida o con espec o al oponen e. Encon amos que los seguido es del candida o
epublicano Donald T ump ue on m´as ac i os en Twi e y m´as belige an es
con a la candida a Hilla y Clin on que ice e sa.
Finalmen e, compa amos nues o conjun o de da os con un es udio simila
en Twi e .
Palab as cla e: an´alisis de sen imien o, Nai e Bayes, Twi e , elecciones,
con incan es, clasi icado es.
Abs ac
This pape conside s he p oblem o ca ego izing ex and use s in he con ex
o social ne wo king, in his case he mic oblogging se ice Twi e . To o his
we downloaded o e 13 million wee s in he days leading up o he 2016 U.S
p esiden ial elec ions.
Using he wee s we downloaded we used a p ocess known as da a cleansing
o build a cohe en da ase . Once he da ase was buil , we used sen imen
analysis o au oma ically assign a label indica ing whe he he wee ’s au ho
suppo s one o he candida es, o on he con a y, ac i ely opposes hem.
Once e e y wee was labeled we compiled a se ies o esul s, he mos in e -
es ing o which being analyzing he beha io o use s who suppo a candida e
while opposing he o he . We ound ha suppo e s o he epublican candida e
Donald T ump we e mo e ac i e on Twi e and mo e bellige en agains he
democ a ic candida e Hilla y Clin on han ice e sa.
Finally, we compa e ou da ase o he da ase o a simila s udy.
Keywo ds: sen imen analysis, sen imen classi ica ion, Nai e Bayes clas-
si ie s, Twi e , elec ions, candida es.
4
Table o con en s
1 In oducci´on 7
2 In oduc ion 9
3 Da a Cap u e 11
3.1 Da aCollec ion............................ 11
3.2 Da aS o age ............................. 12
4 Da a cleaning 15
4.1 Remo e da a o ob ain a cohe en se . . . . . . . . . . . . . . . 15
4.2 Adding in o ma ion . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 Da a quali y dimensions . . . . . . . . . . . . . . . . . . . . . . . 17
4.3.1 Comple eness......................... 18
4.3.2 Uniqueness .......................... 18
4.3.3 Timeliness........................... 18
4.3.4 Validi y ............................ 18
4.3.5 Accu acy ........................... 19
4.3.6 Consis ency.......................... 19
5 Da ase desc ip ion 21
5.1 Use s.................................. 21
5.2 Twee s................................. 22
6 Twee classi ica ion 27
6.1 P ecision and Recall . . . . . . . . . . . . . . . . . . . . . . . . . 30
7 Classi ica ion esul s 33
7.1 Twee esul s ............................. 33
7.2 Use esul s.............................. 37
8 One million ollowe allacy 41
9 Conclusions 43
10 Conclusiones 45
5

6TABLE OF CONTENTS
Chap e 1
In oducci´on
Las edes sociales es ´an cambiando c´omo se compa en no icias e in o maci´on
en la ed. No solo en el campo del pe iodismo, sino muchos o os campos como
po ejemplo la pol´ı ica. Adem´as, pla a o mas como el se icio de mic oblogging
Twi e hacen posible que pe sonalidades pol´ı icas hagan llega sus mensajes a
los usua ios di ec amen e, que a su ez pueden c ea con enido, as´ı omen ando
deba e en la ed.
Nues o abajo se p opone analiza es e deba e en el con ex o de las elec-
ciones gene ales es adounidenses de no iemb e de 2016. Pa a ello desca gamos
du an e la semana que u ie on luga las elecciones odos los wee s que men-
ciona on a alguno de los dos candida os pa a la p esidencia: Donald T ump
(@ ealDonaldT ump) y Hilla y Clin on (@Hilla yClin on). Du an e la semana
de cap u a de da os desca gamos al ededo de 13 millones de wee s.
Como pa e no edosa en el campo nos p oponemos ca ac e iza el ipo de
usua ios que apoya a cada candida o. En pa icula , la no edad de es e abajo
es que nos in e esamos po el sen imien o que los usua ios mues an hacia el
candida o opues o, es deci , al que no apoyan.
An es de pode u iliza es os wee s hicimos una limpieza de da os pa a
ob ene un conjun o cohe en e. A con inuaci´on comp obamos el g ado de cali-
dad del conjun o median e un p oceso de audi o ´ıa desc i o en el cap´ı ulo 4.
El an´alisis de sen imien o es una ´ecnica com´un que se ha u ilizado en muchos
abajos pa a ob ene y clasi ica au om´a icamen e el signi icado de ex os. Es a
´ecnica se ha u ilizado en edes sociales[1], en el con ex o come cial pa a analiza
opiniones de usua ios, y, como en nues o caso, pa a es udia las opiniones de
los usua ios de Twi e en elecciones pol´ı icas[2].
Es e abajo con empla el p oblema de clasi ica au om´a icamen e millones
de wee s u ilizando un clasi icado p obabil´ıs ico conocido como clasi icado
Bayesiano ingenuo. Es os clasi icado es pueden emplea se pa a oma deci-
siones bina ias simples, como po ejemplo deci nos si un nomb e es masculino
o emenino, o si un ex o habla de un p oduc o de mane a posi i a o nega i a.
Nues o caso es m´as complicado ya que un wee puede ene sen imien o
nega i o hacia T ump y posi i o hacia Clin on o ene sen imien o neu al hacia
7
8CHAPTER 1. INTRODUCCI ´
ON
ambos. Es o da luga a muchas posibilidades y es un p oblema que a on amos
en el cap´ı ulo 6. Pa a minimiza el n´ume o de posibilidades de clasi icaci´on
abo damos el p oblema de o ma pa alela. Es deci , clasi icamos cada wee
una ez po candida o.
Una ez clasi icados odos los wee s podemos p ocede a analiza los y ex-
ae esul ados aplicados a wee s y usua ios como e emos en el cap´ı ulo 7.
Pa a aplica los esul ados a los usua ios expo amos la clasi icaci´on de wee s
pa a clasi ica los po usua ios. Es os esul ados se encuen an en la secci´on 7.2.
Una ez e minado el p oceso encon amos unas di e encias signi ica i as
en e los seguido es de ambos candida os y e emos que en la mayo pa e de
los casos, los n´ume os bene ician al candida o Donald T ump.
Chap e 2
In oduc ion
Social ne wo king is changing how news and in o ma ion is sha ed online. No
only in jou nalism, bu in ields like ma ke ing and poli ics as well. Social
ne wo king also gi es poli icians a pla o m ha allows hem o di ec ly in e ac
wi h use s, who a he same ime c ea e con en , leading o online deba e.
This pape analyzes his deba e in he con ex o he No embe 2016 U.S
elec ions. In he week leading o he elec ions we downloaded o e 13 million
wee s published by use s who men ioned one o he wo p esiden ial candida es:
Donald T ump (@ ealDonaldT ump) and Hilla y Clin on (@Hilla yClin on).
P e ious wo ks ha e s udied Twi e bu ew ha e ied o cha ac e ize he
ype o use s who suppo each candida e. The no el y o his pape is ha
we examine he sen imen ha use s show owa d he candida e hey do no
suppo .
The wee s we used could no be used in hei aw o ma and we used a
echnique known as da a cleansing o build a cohe en da ase . We ollowed his
up by assessing he quali y o he da a, a p ocess desc ibed in chap e 4.
Sen imen analysis is a me hod ha has been equen ly used in o he s udies
o au oma ically ob ain and classi y he meaning o ex s. This me hod has
been used in social ne wo king[1], and, like in ou case, o s udy he opinions o
Twi e use s du ing poli ical elec ions[2].
Once he cohe en da ase was buil we encoun e ed he p oblem o classi y-
ing e e y wee . This was done using a p obabilis ic classi ie known as a Nai e
Bayes Classi ie . These classi ie s ca be used o make simple bina y decisions.
Fo example, we could use a classi ie o de e mine whe he a name is male o
emale, o i a p oduc e iew is posi i e o nega i e.
Ou case is mo e complex because a wee can ha e nega i e sen imen o-
wa d T ump and posi i e sen imen owa d Clin on, o a wee could simply be
s a ing a ac and no ha e sen imen a all. The numbe o classi ying choices
makes he decision ha de o make and is mo e ho oughly discussed in chap e
6. We made he classi ie ’s job easie by aking a pa allel app oach, meaning
each wee was classi ied one ime o each candida e.
When e e y wee in he da ase was classi ied we we e able o su ey hem
9
16 CHAPTER 4. DATA CLEANING
C ea e wee s collec ion
C ea e use s collec ion
Remo e duplica es and wee s published by suspended o dele ed accoun s
Remo e e wee s whose o iginal wee is ou side o he ime ame
Remo e all wee s wi hou a sou ce. 2.44 million wee s we e emo ed
Figu e 4.1: Da a emo al cleanup
4.2 Adding in o ma ion
To ha e a be e unde s anding o he da ase we also added addi ional ields
o bo h use and wee documen s. We added h ee ields ela ed o e wee s o
e e y wee documen :
RT sc een o wee documen s. Indica es he o iginal use ’s Twi e handle.
I s alue is no hing(””) i he wee was ne e e wee ed
sou ce he o iginal wee ’s ID. I s alue is no hing(””) i he wee was ne e
e wee ed
RTin numbe o imes he wee has been e wee ed. 0 i he wee was ne e
e wee ed
We also compiled a summa y o each use :
wee s : he use ’s Twi e ac i i y epo
all : use ’s wee coun a he ime o hei i s wee in he ime ame
o al : use ’s wee coun in he ime ame. o al = RT + o iginal
RT : numbe o e wee s in he ime ame
o iginal : numbe o o iginal wee s in he ime ame
Finally, ields ela ed o sen imen analysis we e added du ing he classi ica-
ion p ocess. These ields will be discussed in chap e 4: Twee classi ica ion.

4.3. DATA QUALITY DIMENSIONS 17
A de ailed illus a ion o he s eps aken o add in o ma ion o he da ase
can be ound in igu e 4.2.
Add e wee in o ma ion o wee s and use s
Calcula e and add ”men ions” ield o each use
Calcula e and add ”RTin” ield and wee s pos ed by each use
Add sen imen analysis de ails o use s and wee s
Figu e 4.2: Da a addi ion lowcha
4.3 Da a quali y dimensions
Once we cleaned he da ase and added new ields o he documen s, we ex-
amined he da ase o measu e he quali y o he da a. To do his, we based
ou analysis on a pape i led The Six P ima y Dimensions o Da a Quali y
Assessmen [3]. Ano he in e es ing book on da a quali y is Da a Quali y As-
sessmen [4]. This pape ou lines six key ‘dimensions’ ecommended o be used
when assessing o desc ibing da a quali y. In his con ex , he au ho s ake he
e m da a quali y dimension o mean: some hing ha can ei he be measu ed
o assessed in o de o unde s and he quali y o he da a. The six dimensions
a e:
•Comple eness
•Uniqueness
•Timeliness
•Validi y
•Accu acy
•Consis ency
18 CHAPTER 4. DATA CLEANING
4.3.1 Comple eness
Comple eness is de ined as The p opo ion o s o ed da a agains he po en ial
o ”100% comple e”. In ou case, his would be calcula ed by di iding he o al
numbe o wee s and use s in ou da abase by 100% o he wee s published in
he ime ame and he ac i e use s.
Du ing he da a cap u e phase ou collec ion p og am expe ienced se e al
sho pauses du ing which we we en’ able o collec da a. Because ou p og am
was no unning, we a e unable o know how many wee s we e missed, and hus
unable o accu a ely measu e comple eness. Howe e , we es ima e ha he da a
we missed is a small pe cen age o ou da ase , and ha he missed da a is no
c i ical.
A way we can calcula e he comple eness o ou da ase is by compa ing he
ini ial da a o he inal usable da ase .
Comple eness =7.803.405
13.358.219 ×100 = 58.4%
4.3.2 Uniqueness
Uniqueness is de ined as No hing will be eco ded mo e han once based upon
how ha hing is iden i ied. One way o calcula ing he deg ee o uniqueness
would be o di ide he numbe o unique documen s by he numbe o o al
documen s. Ou case is special because we s o e o iginal wee s and e wee s.
I we we e o conside e wee s exac copies o wee s, he uniqueness o he
da ase would be o 71.1%.
Howe e , in ou case we conside each e wee o be unique because we don’
only s o e he wee ’s ex , bu da a abou he use s who a e o wa ding he
wee s. Wi h ha in mind, we can say ha ou da ase ’s uniqueness is o 100%.
4.3.3 Timeliness
Timeliness is de ined as The deg ee o which da a ep esen eali y om he
equi ed poin in ime. This can be unde s ood as he ime ha has elapsed
be ween da a collec ion and da a s o age. Ou da a collec ion p og am builds
and s o es he MongoDB documen s as soon as hey a e de ec ed, so he e is a
high deg ee o imeliness.
4.3.4 Validi y
Validi y is de ined as Da a a e alid i i con o ms o he syn ax ( o ma , ype,
ange) o i s de ini ion. In ou case alidy means ha ou da a con o ms o a
se ies o ules. One o he mos impo an aspec s o ou da ase a e ha da es
and imes a e s o ed in he same imezone. Because o his, all o ou da a is
s o ed using Zulu ime and da e.
4.3. DATA QUALITY DIMENSIONS 19
4.3.5 Accu acy
Accu acy is de ined as The deg ee o which da a co ec ly desc ibes he ” eal
wo ld” objec o e en being desc ibed. We can apply his measu e o ou da ase
by asking one ques ion: how closely do ou wee documen s esemble he wee s
hemsel es in he ” eal wo ld”?. We’ e able o s o e only wha he Twi e API
p o ides us. In some cases o long wee s, Twi e didn’ p o ide he wee ’s
ull ex . Ins ead, i ’s cu o and eplaced by an ellipsis (...). We belie e he
es o ou da a o be accu a e.
4.3.6 Consis ency
Consis ency is de ined as The absence o di e ence, when compa ing wo o
mo e ep esen a ions o a hing agains a de ini ion. The uni o measu e is pe -
cen age and i can be ob ained by compa ing he numbe o inaccu a e eco ds
agains he o al numbe o eco ds. An in e es ing cha ac e is ic abou his
dimension is ha consis ency can be achie ed wi hou alidi y o accu acy. I
we apply his measu e o ou da ase we can say ha i ’s consis en because
o he way he da abase was buil . I a use ’s documen speci ies hey ha e
published en wee s, en dis inc wee s will appea in he wee collec ion by
ha use . This makes ou da ase 100% consis en .
Wi h hese da a quali y dimensions in mind, we eel con iden o he quali y
o ou da ase .
20 CHAPTER 4. DATA CLEANING
Chap e 5
Da ase desc ip ion
To ha e a be e unde s anding o he da a ob ained we que ied he da ase and
compiled a numbe o s a is ics ha will be illus a ed in his chap e . The e
will be a subsec ion o he wo a o emen ioned en i ies: use s and wee s.
5.1 Use s
To compa e he use s wee ing each candida e we sepa a ed use s by g ouping
hem oge he depending on which candida e hey men ioned: @Hilla yClin on
o @ ealDonaldT ump.
The ob ious i s s ep is o see wha pe cen age o use s men ioned each can-
dida e in hei wee s. Bo h candida es we e men ioned equally in ou da ase ;
no disce nible di e ence was obse ed.
Clin on
T ump 51.2
48.8
Men ions (pe cen age)
Ve i ied accoun s a e hose ha Twi e conside s o be o public in e es .
The owne s o hese accoun s a e usually news ne wo ks, celeb i ies, pe o me s,
use s specializing in key in e es a eas, and o he s. Twi e uses a blue e i ied
badge o le o he s know ha an accoun o public in e es is au hen ic.
I is no easy o ob ain a e i ied badge; he as majo i y (98.65%) o use s in
ou da ase a e un e i ied use s. Accoun s mus mee a a ie y o equi emen s
o hem o be e i ied, he mos impo an being ha all published wee s by he
accoun mus ha e an open p i acy se ing, meaning wee s by he accoun a e
able o be seen by any Twi e use . Addi ionally, he owne o he accoun mus
21

22 CHAPTER 5. DATASET DESCRIPTION
ill ou a o m and submi i o Twi e , who assesses each eques indi idually
and app o es o denies i . I he eques is denied, use s mus wai 30 days
be o e submi ing ano he .
0 20 40 60 80 100
1.3598.65
Pe cen age o e i ied accoun s in blue (da ase )
Since e i ied accoun s a e deemed o be o public in e es , hey a ac mo e
use s and he e o e ha e a a la ge a e age amoun o ollowe s han egula
accoun s do. The a e age numbe o ollowe s o e i ied accoun s is 173054
whe eas un e i ied accoun s ha e an a e age o 1387 ollowe s.
I we pe o m he same calcula ion on each candida e’s se we saw a consid-
e able di e ence:
0 20 40 60 80 100
3.2296.79
Pe cen age o e i ied accoun s in blue (Clin on)
0 20 40 60 80 100
8.6791.33
Pe cen age o e i ied accoun s in blue (T ump)
This di e ence is impo an because, as we’ e seen, e i ied use s a e mo e
in luen ial han s anda d use s and in his case i could be said ha wee s
men ioning T ump had mo e exposu e han wee s men ioning Clin on.
5.2 Twee s
A e cleaning up he da a we had a sizable da ase o wee s ha we di ided
in o wo g oups in he same manne as we di ided he use s: using men ions.
These g oups a e no disjoin ; a wee ha men ions bo h candida es will be
pa o bo h g oups.
5.2. TWEETS 23
Clin on
T ump 53.75
46.25
Pe cen age o wee s men ioning each candida e
An impo an aspec abou ou da ase o wee s is how many o he wee s
a e ac ually o iginal and how many wee s a e simply e wee s, o o wa ded
wee s.
O iginal
Re wee 63.2
36.8
Pe cen age o wee s s e wee s (da ase )
We ound ha du ing he ime we buil he da ase , use s in e ac ed wi h each
o he by o wa ding o he s’ wee s, a he han pos ing hei own o iginal wee s.
We also pe o med he same calcula ion on each candida e’s wee da ase and
no signi ican di e ence was obse ed. Indeed, he esul s a e almos iden ical
o he da ase o all wee s.
We can di ide he o iginal wee s in he da ase in wo ca ego ies: o iginal
wee s wi h e wee s and o iginal wee s ha we en’ e wee ed.
Re wee ed
No e wee ed 85.1
14.9
Pe cen age o e wee ed o iginal wee s
O ou da ase o se e al million wee s, only 36.8% o wee s a e o iginal wee s.
Only 14.9% o hose o iginal wee s we e e wee ed. This means ha 63.2% o
he wee s in he da ase a e ac ually e wee s o a small 5.49% o he en i e
da ase . This leads us o belie e ha a as majo i y o wee s all on dea ea s.
The mos e wee ed wee men ions bo h candida es, so i ops he lis o
he mos e wee ed wee o bo h candida e’s da ase . The au ho o his wee
is @ladygaga, he second mos ollowed accoun on Twi e and he wee was
24 CHAPTER 5. DATASET DESCRIPTION
o wa ded 17993 imes.
The mos e wee ed wee by an un e i ied au ho is pa o @ ealDon-
aldT ump’s da ase , and was e wee ed 17143 imes:
Con e sely, he mos e wee ed wee o @Hilla yClin on’s da ase by an un e -
i ied au ho was e wee ed 9546 imes:
5.2. TWEETS 25
The a e age numbe o e wee s o a wee wi h a leas one e wee is 11.5
e wee s. I we compa e he candida e’s da ase we can see ha e wee ed
wee s men ioning @ ealDonaldT ump a e o wa ded mo e imes han hose
men ioning @Hilla yClin on.
Clin on
T ump 12.06
10.64
A e age numbe o e wee s o a e wee ed wee
Wi h he da a a hand we we e able o plo he equency o wee s by ime
and da e shown in igu e 3.1. We can clea ly see he equency o ac i i y ise
du ing he day and all a nigh , as well as a conside able su ge in Twi e
ac i i y a e elec ion day (No embe 8) once he esul s came in.
Figu e 5.1: Thousands o wee s pe day and hou
32 CHAPTER 6. TWEET CLASSIFICATION
Label Recall
TPos 85.68%
Ze oT 89.81%
TNeg 86.9%
Figu e 6.5: Recall esul s o he epublican candida e Donald T ump
Label Recall
HPos 91%
Ze oH 97%
HNeg 58%
Figu e 6.6: Recall esul s o he democ a ic candida e Hilla y Clin on
Candida e O e all accu acy
Donald T ump (R) 87.8%
Hilla y Clin on (D) 86.4%
Figu e 6.7: O e all accu acy esul s o bo h candida es

Chap e 7
Classi ica ion esul s
The classi ie we desc ibed in he p e ious chap e assigned one o mo e labels o
each wee depending on i s sen imen . Once he classi ie was inished labeling
wee s, we had a se o 7,8 million labeled wee s and 1,5 million use s o analyze.
7.1 Twee esul s
A e he classi ie was inished, h ee new ields we e added o e e y wee in
he da ase :
opinion : i he wee has sen imen i ’s equal o 1, o he wise i ’s 0
hlabel : i he sen imen is posi i e owa d Clin on he ield akes a alue o 1.
I i ’s nega i e i will be -1, and i i has no sen imen owa d Clin on i will
be 0
label : i he sen imen is posi i e owa d T ump he ield akes a alue o 1.
I i ’s nega i e i will be -1, and i i has no sen imen owa d Clin on i will
be 0
The i s s ep he classi ie ook, as men ioned in chap e 4, was o decide i
a wee has sen imen o no . This esul s in wo disjoin se s: one se o wee s
wi h asce ainable sen imen and a second se o wee s wi hou sen imen .
Wi hou sen imen
Wi h sen imen 62.4
37.6
The classi ie p o ided us wi h a se o 4,86 million wee s wi h sen imen , 62.4%
o he o iginal se . This is he se ha we will conside in his chap e . The se
o wee s wi hou sen imen will no be discussed any u he .
33
34 CHAPTER 7. CLASSIFICATION RESULTS
The ollowing g aph illus a es he numbe o wee s o each indi idual la-
bel, ha is o say, only one label and no o he , in housands o wee s.
TPos
HPos
TNeg
HNeg 948.1
1,020.71
966.75
1,386.42
Label appea ances ( housands o wee s)
As we can see, he e a e clea ly mo e wee s wi h posi i e sen imen owa d
T ump han Clin on, and he p opo ion o wee s wi h posi i e sen imen s
nega i e sen imen is la ge o T ump han o Clin on. Indeed, Clin on’s a io
is almos 1:1, which means ha o e e y wee wi h posi i e sen imen , he e
is ano he wi h nega i e sen imen .
Now we can ake a look a he wee s ha we e assigned mo e han one
label. Two combina ions will no appea because hey’ e complemen a y: TPos
and TNeg, HPos and HNeg.
TNeg & HNeg
TNeg & HPos
TPos & HNeg
TPos & HPos 58.13
309.76
101.82
77.06
Label combina ions ( housands o wee s)
In his case one o he combina ions immedia ely cap u es ou a en ion.
The numbe o wee s wi h a posi i e sen imen owa d T ump and nega i e
sen imen owa d Clin on mo e han iples he second mos obse ed combi-
na ion o labels. We in e p e his o mean ha T ump ollowe s, hose who
publish wee s wi h posi i e sen imen owa d T ump, a e a mo e likely o
use Twi e o speak ill o Clin on han Clin on suppo e s a e o speak ill o
T ump.
7.1. TWEET RESULTS 35
Pe haps unsu p isingly, he leas obse ed combina ion o labels was he com-
bina ion o posi i e sen imen owa d bo h candida es.
We que ied ou da ase o ind he mos ep esen a i e wee o each la-
bel, which we de e mined o be he one wi h he mos e wee s and a co ec ly
guessed sen imen by he classi ie . We will begin by inding he mos ep esen-
a i e wee s ha we e only assigned one label. The wee s we e aken di ec ly
om Twi e in o de o be e illus a e hem.
We’ll also show how many imes he wee was e wee ed a he ime we in-
ished cap u ing ou da a. I we we e o sea ch o he wee oday, he numbe
o e wee s would di e because he wee con inued o be e wee ed a e we
inished he p ocess o cap u ing da a.
The mos ep esen a i e wee o he label TPos was pos ed by he use
@T ump bi d and i has 7705 e wee s:
The mos ep esen a i e wee o he label HPos was pos ed by he use
@Vigilan eA is and i has 16485 e wee s:
The mos ep esen a i e wee o he label TNeg was pos ed by he use
@ladygaga and i has 17993 e wee s. This is he same wee we saw in he
second sec ion o he hi d chap e :
36 CHAPTER 7. CLASSIFICATION RESULTS
The mos ep esen a i e wee o he label HNeg was pos ed by he use
@TeamT ump and i has 11462 e wee s:
Now we can que y he se o ind he mos ep esen a i e wee s ha he
classi ie assigned mo e han one label o.
In he case o he combina ion o he labels TNeg and HNeg he mos ep-
esen a i e wee was pos ed by he use @joe012594 and was e wee ed 8882
imes:
In he case o he combina ion o he labels TNeg and HPos he mos ep esen-
a i e wee was pos ed by he use @syeddoha and was e wee ed 2845 imes:
7.2. USER RESULTS 37
In he case o he combina ion o he labels TPos and HNeg, he mos epea ed
combina ion o labels, he mos ep esen a i e wee was pos ed by he use
@JoseANunez1 and was e wee ed 4555 imes:
Finally, in he case o he combina ion o he labels TPos and HPos, he
leas equen combina ion o labels, he mos ep esen a i e wee was pos ed
by he use @Ha lan and was e wee ed 3230 imes:
7.2 Use esul s
Once e e y wee was labeled, a sen imen summa y o he e e y use ’s published
wee s was added o he use documen s:
h0 : numbe o wee s published du ing he ime ame wi h neu al sen imen
owa d Clin on
hp : numbe o wee s published du ing he ime ame wi h posi i e sen imen
owa d Clin on
hn : numbe o wee s published du ing he ime ame wi h nega i e sen imen
owa d Clin on
0 : numbe o wee s published du ing he ime ame wi h neu al sen imen
owa d T ump

38 CHAPTER 7. CLASSIFICATION RESULTS
p : numbe o wee s published du ing he ime ame wi h posi i e sen imen
owa d T ump
n : numbe o wee s published du ing he ime ame wi h neu al sen imen
owa d T ump
ph0 : h0/ wee s. o al
php : hp/ wee s. o al
phn : hn/ wee s. o al
p 0 : 0/ wee s. o al
p p : p/ wee s. o al
p n : n/ wee s. o al
Whe e he ollowing is ue:
0+ p+ n = wee s. o al = wee s.RT + wee s.o iginal
h0+hp+hn = wee s. o al = wee s.RT + wee s.o iginal
Using hese ields we used a o mula using a 20% ma gin o e o o asce -
ain whe he a use suppo s a candida e. I we we e o ind T ump suppo e s
we would use his o mula o ind use s ha make i ue:
p p >(p 0 + p n + 0.2) & wee s. o al >5
Con e sely we can use he same o mula o ind use s who ac i ely oppose
T ump:
p n >(p 0 + p p + 0.2) & wee s. o al >5
The o al amoun o suppo e s is p opo ionally low o he o al numbe o
use s. Howe e , aking in o accoun he ma gin o e o o he classi ie and he
s ic ness o ou measu e we can be easonably su e ha hese use s ha e been
classi ied co ec ly.
Using hese measu es we we e able o build ou subse s o use s:
T ump suppo e s
Clin on suppo e s
Agains T ump
Agains Hilla y 5,061
3,944
6,859
9,775
Numbe o use s by sen imen
The esul s seem o con i m he da a obse ed in he p e ious sec ion: he e
7.2. USER RESULTS 39
a e mo e T ump suppo e s in ou da ase and mo e use s a e publishing wee s
wi h nega i e sen imen owa d Clin on han owa d T ump.
Using hese ou new subse s we can ake a look a hei cha ac e is ics and
compa e he ypes o use :
T ump suppo e s
Clin on suppo e s
Agains T ump
Agains Hilla y 22.28
13.57
13.62
24.22
A e age numbe o published wee s by use ype
Again, his measu emen seems o be in line wi h wha we’ e obse ed so a :
T ump has mo e suppo e s han opponen s in ou da ase , and he opposi e
is ue o Clin on. Fu he mo e, T ump suppo e s publish conside ably mo e
wee s han Clin on suppo e s.
T ump suppo e s
Clin on suppo e s
Agains T ump
Agains Hilla y 0.18
3.06
2.85
0.61
Pe cen age o e i ied use s
An in e es ing phenomenon akes place i we que y he da ase o ob ain he
pe cen age o e i ied use s o he di e en ypes o use . While e i ied use s
a e a small ac ion o he da ase we can see ha he esul s di e om he
”p o T ump” end: Clin on suppo e s and T ump opponen s show a bigge
p opo ion o e i ied use s han T ump suppo e s and Clin on opponen s.
A possible explana ion o his is ha T ump suppo e s a e mo e ac i e
on Twi e and ha e mo e p esence han Clin on suppo e s, while Clin on
suppo e s a e no as ocal bu hei wee s ha e mo e public in e es han
wee s published by T ump suppo e s.
40 CHAPTER 7. CLASSIFICATION RESULTS
Now ha we ha e iden i ied T ump and Clin on suppo e s we can que y he
da ase o ind use s who suppo a candida e and ac i ely oppose he o he .
We conside a suppo e o oppose he o he candida e i a leas 20% o hei
wee s ha e nega i e sen imen owa d he o he candida e.
T ump suppo e s agains Clin on
Clin on suppo e s agains T ump 12
51
Pe cen age o suppo e s ha oppose he o he candida e
The esul s a e clea and hey con i m wha we suspec ed: T ump suppo -
e s a e a mo e likely o publish nega i e wee s abou Clin on han Clin on
suppo e s a e o publish nega i e wee s abou T ump.
Chap e 8
One million ollowe allacy
Du ing he cou se o ou esea ch we encoun e ed a e m coined million ollowe
allacy[6] by A ni (A ni 2009) who claims ha ha ing a la ge amoun o
ollowe s does no always equa e o being in luen ial in he Twi e wo ld. We
based ou analysis on he pape [7] i led Measu ing Use In luence in Twi e :
The Million Followe Fallacy ha discusses and expands upon he o iginal A ni
pape .
Fi s , we mus de ine wha ha ing in luence on Twi e means. The Cam-
b idge Dic iona y de ines in luence as ” he powe o ha e an e ec on people
o hings, o someone o some hing ha ing such powe ”. Since he e is no con-
c e e way o measu ing in luence on Twi e , we ollow Cha e al’s example and
unde line h ee ac i i ies ha we will use o measu e in luence:
•Followe s: he numbe o ollowe s he use has di ec ly indica es he size o
hei audience.
•Men ions: he numbe o imes he use has been men ioned indica es he
use ’s means o in e ac wi h o he Twi e use s
•RTin: he numbe o imes he use ’s o iginal wee s ha e been sha ed indi-
ca es he use ’s abili y o gene a e con en
In o de o in es iga e how he h ee measu es co ela e, we compa ed he el-
a i e in luence o he op 2033 use s based on numbe o ollowe s.
The ollowing able ep esen s a g aphical example o he da ase ’s op 10
use s based on numbe o ollowe s:
41