Measuring polarization on social media

measuringpolarizationonsocialmediamichael mathioudakis

computerscienceaalto university

http://www.michalis.co

2

socialmedia

MichaelMathioudakis

consumecontentnewsaboutfriends,politics,favoriteartists

generatecontentshareexperiences,interestingarticles

interactwithotherscomment,rate,anddiscuss

hundredsofmillionsofactiveusers

whystudysocialmedia

awindowintothethoughtsandactionsofpeople

howpeoplespendtheirtimewhatattractstheirattention

whattheythinkaboutpoliticalissueshowtheyinteractsocially

uniqueopportunityforsocialscientistsstudyhumanbehavior

atlargescale&finedetail

MichaelMathioudakis 3

butyouareacomputerscientistdevelopalgorithmsto

extractinsightsfromsocialmedia

twochallenges1- volumeofdata

2- complexityofsocialbehaviorcallsforcomplexmodels (e.g.,graphs)

thatrequirefastalgorithms

solutionalgorithmicmethods,softwaretools andsystems

forlarge-scalecomputationanalyze efficiently&automatically largeamountsofdata

MichaelMathioudakis 4

5

noinformationbarriers

citizenjournalism

socialconnectivity

…

…

harassment

fakenews

polarization

socialmedia:goodandbadsides

MichaelMathioudakis

polarization

political orsocialpolarization‘theactofseparatingormakingpeopleseparateinto

twogroupswithcompletelyoppositeopinions’*

relatedterm:controversy‘publicdiscussionandargumentaboutsomethingthat

manypeoplestronglydisagreeabout’*

*oxfordenglish dictionary

MichaelMathioudakis 6

whystudypolarization?

polarizationcanbelinkedtoadverseeffects

socialsegmentationstereotypes

echochambers

goalunderstandandmitigate them

MichaelMathioudakis 7

8

0102030405060708090100

DonaldTrumpApprovalRating%February2017

Democrats Republicans

MichaelMathioudakis

9

ideologyofUSApublicsource:PewResearchCenter

1994

2014

DemocratsandRepublicanshavedriftedapart

MichaelMathioudakis

whystudypolarizationonsocialmedia

extractinsightsonthepolarizationprocess

somespecificquestionswhatarethepolarizingissues?

doespolarizationincreaseovertime?wesawonesuchinstance:USpublic

howcouldwe‘nudge’peopleawayfromextremepolarization?

ispolarizationlinkedtoechochambers?dosocialmediaincreasepolarization?

MichaelMathioudakis 10

this

talk

inthistalk…

twitter

algorithmstomeasureit

long-termstudy

nextsteps

11

polarization

MichaelMathioudakis

12

retweetstweet

replies

connections

interactionsMichaelMathioudakis

twitter

microbloggingplatformsince2006;300millionactiveusers

userspostshortmessages- ‘tweets’

13MichaelMathioudakis

globalstructureofinteractions

nodes

14

useraccounts

edges retweetsrepliesconnections

interactiongraph

MichaelMathioudakis

1%sampleofalltweetsfromtheinternetarchive

data

inwhatfollows…

15MichaelMathioudakis

twitteralgorithmstomeasureit

long-termstudynextsteps

polarization

QuantifyingControversyinSocialMedia. WSDM2016.WithKiranGarimella,Gianmarco DeFrancisci Morales,AristidesGionis.

howpolarizedisadiscussion?

e.g.,obamacare inJanuary2017

16MichaelMathioudakis

howcouldweapproachthisquestion?

content?doopposingsidessaydifferentthings?

sentiment?dopolarizedtopicsexhibitwider

rangeofemotions?

interactions?dopeopleinteractmorewiththeirownside?

MichaelMathioudakis 17

let’strythis

methodtemplate

buildaninteractiongraphseveraltypestotry

retweets,replies,connections

istheinteraction graph polarized?

output:polarizationscore

18

notpolarized polarizedtwosides,wellseparated

MichaelMathioudakis

pipeline

19MichaelMathioudakis

topic

evaluatio

n

Graph Building

Graph Partitioning

Controversy Measure

Polarization

whattypeofinteractiongraphshouldweuse?

howtofindtwosidesinthe

graph?

howtomeasuretheseparationbetween

twosides?

doweidentifypolarized

discussions?

retweetsreplies

connections

20MichaelMathioudakis

topic

evaluatio

n

Graph Building

Graph Partitioning

Controversy Measure

Polarization

whattypeofinteractiongraphshouldweuse?

howtofindtwosidesinthe

graph?

howtomeasuretheseparationbetween

twosides?

doweidentifypolarized

discussions?

21

partitionintotwosides

MichaelMathioudakis

topic

evaluatio

n

Graph Building

Graph Partitioning

Controversy Measure

Polarization

whattypeofinteractiongraphshouldweuse?

howtofindtwosidesinthe

graph?

howtomeasuretheseparationbetween

twosides?

doweidentifypolarized

discussions?

partitioning

manyexistingalgorithms

spectralclusteringlabelpropagationmetis

MichaelMathioudakis 22

wellconnected

badlyconnected

randomwalkcontroversy(RWC)

23MichaelMathioudakis

topic

evaluatio

n

Graph Building

Graph Partitioning

Controversy Measure

Polarization

whattypeofinteractiongraphshouldweuse?

howtofindtwosidesinthe

graph?

howtomeasuretheseparationbetween

twosides?

doweidentifypolarized

discussions?

randomwalkcontroversy(RWC)

polarizationscorehowdifficulttoreachc fromB

24

A andB:twosides

c andd:centralnodes

andvalues-1to1

themoredifficultthehigherthescorehowdifficulttoreachdfromA

MichaelMathioudakis

cd

A B

randomwalk

randomwalk

25MichaelMathioudakis

discreteprocess onagraphstep1,2,3,…

ateachstep,weareatonenodeatnextstep,wemovetoanearbynode

- followinganedge atrandom

specifywherethewalkstarts (e.g.,atc)wherethewalkends (e.g.,atd)

cd

randomwalkcontroversy(RWC)

26

A andB:twosides

c andd:centralnodes

MichaelMathioudakis

cd

A B

randomwalkstart:randomnodeofA/B50%chancethatwedoeitherend:atcentralnodescofsideA ordofsideB

PXY =P(startedinX |endedinY)

considertwoindependentinstancesoneending atthecentralnodeofsideA,theotheratthecentralnodeofsideB

RWC=PAAPBB - PBAPAB

bothrandomwalksstartedinthesidewheretheyended

bothrandomwalksstartedinasideotherthanwheretheyended

values?

randomwalkcontroversy(RWC)

27MichaelMathioudakis

topic

evaluatio

n

Graph Building

Graph Partitioning

Controversy Measure

Polarization

whattypeofinteractiongraphshouldweuse?

howtofindtwosidesinthe

graph?

howtomeasuretheseparationbetween

twosides?

doweidentifypolarized

discussions?

28MichaelMathioudakis

topic

evaluatio

n

Graph Building

Graph Partitioning

Controversy Measure

Polarization

whattypeofinteractiongraphshouldweuse?

howtofindtwosidesinthe

graph?

howtomeasuretheseparationbetween

twosides?

doweidentifypolarized

discussions?

evaluation

non-polarized topics

indian beefban,nemtsov protests,netanyahu UScongressspeech,

baltimore riots,ukrainepolarizedtopics

germanwings planecrash,sxsw,mother’sday,jurassic worldmovie,

nationalkissingday

29MichaelMathioudakis

populartopicsof2015graphswith1– 150thousandnodes

examples:

Graph Building

Graph Partitioning

Controversy Measure

Polarization

results

MichaelMathioudakis 30

doesthepipelinedistinguishpolarizedfromnon-polarizedtopics?

retweetsreplies

connections

twosides randomwalkcontroversy(RWC)

topic

(ifweuseretweets)yes!

results

nemtsovprotests

indianbeefban

sxswconference

germanwingsplanecrash

interactiongraphs:retweets31MichaelMathioudakis

polarized

topics

non-polarizedtopicshighRWC lowRWC

results

32

retweets replies

interactiongraphsfornemtsov protests

MichaelMathioudakis

analgorithmicwaytoquantifypolarization

MichaelMathioudakis 33

retweets twosides randomwalkcontroversy(RWC)

topic

basedonstructureofinteractionslanguage-independent

Graph Building

Graph Partitioning

Controversy Measure

Polarization

newmethodcanbedeployedinthewild

inwhatfollows…

34MichaelMathioudakis

twitteralgorithmstomeasureit

long-termstudynextsteps

polarization

Graph Building

Graph Partitioning

Controversy Measure

TheEbbandFlowofControversialDebatesonSocialMedia.ICWSM2017.TheEffectofCollectiveAttentiononControversialDebatesonSocialMedia.WebScience 2017.WithKiranGarimella,Gianmarco DeFrancisci Morales,AristidesGionis.

Polarization

polarizationovertime

haspolarizationaroundcontroversialtopicsincreasedovertime?

35MichaelMathioudakis

polarizationovertime

data1%sampleofalltweetsfromtheinternetarchive

September2011toSeptember2016

methodforagiventopic(e.g.,obamacare)

retrieverelatedtweetsbuildaninteractiongraphforeachday

usingretweetsmeasureRWCscore

MichaelMathioudakis 36

RWCvsTime

37MichaelMathioudakis

September2011

September2016

volumeofactivityspikesatmajorevents

38MichaelMathioudakis

39MichaelMathioudakis

doespolarizationspikewithvolume?measurerwc vsvolumeofactivity

40MichaelMathioudakis

RWCvsVolume

highervolume

higherco

ntroversy

41MichaelMathioudakis

highervolume highervolume

RWCvsVolume

MichaelMathioudakis 42

highervolume

higherco

ntroversy

rwc

summary

43MichaelMathioudakis

twitteralgorithmstomeasureit

long-termstudynextsteps

polarization

whystudypolarizationonsocialmedia

extractinsightsonthepolarizationprocess

somespecificquestionswhatarethepolarizingissues?

doespolarizationincreaseovertime?wesawonesuchinstance:USpublic

howcouldwe‘nudge’peopleawayfromextremepolarization?

ispolarizationlinkedtoechochambers?dosocialmediaincreasepolarization?

MichaelMathioudakis 44

this

talk

nextsteps

thank you!

45MichaelMathioudakis