human behavior as recorded on the web webst symposium thursday, february 24 th, 2011 imperial palace...

Human Behavior as Recorded on the Web

WebST SymposiumThursday, February 24th, 2011Imperial Palace Hotel, Seoul

Sue Moon

Graduate Program of Web Science and Technology &Department of Computer Science

KAIST

2

Historic Records of Human Activities

함부라비 법전 ?

알타미라 동굴 그림

훈민정음 해례

3

Personal Correspondents

4

Come Internet

• Your explicit trace of existence– Emails– Chat room activities– Messenger activities– Files you create/modify/delete– Newsgroup– Comments– Web

• Your implicit trace– Search keyword logs

5

MyLifeBits

Picture of LifeBits (MSR Mountain View guy)

6

In the Middle East

7

Information Diffusion

• Reflects “potentials of power transition”• Egypt, Libya, MENA ( 뭐의 약자 ?)• Twitter/FB critical or supplementary?– One thing for sure: records of word-of-mouth

spreading

8

Traces of Twitpic

• Guaranteed single source• Unique URL• Twitter-internal starting point

9

Some Twitter specifics

• Our unit of information = Twitpic– Affiliated web site of pictures

• Why not tweets themselves?– We are looking at tweets of trending topics

• Why not general URLs?– Typically in shortened forms (bit.ly, tinyurl, t.co)– Can be in multiple shortened forms– Hard to identify sources

Duration of Twitpic Spreading

# of Tweets

median duration

(day)

Which Twitpic is most popular?

• # of tweeted– Form of recommendation (quality)

• # of viewed– User clicks on URL (popularity)

• # of total followers– Measure of Information exposure

# of tweeted

# of viewed

# of followers

Short-Lived, Ephemeral Fame# of followers# of tweeted

15

Case Study : Evolution of #Views

• # of user clicks on the Twitpic URL– limitation : some Twitter clients show the photos

without clicks (no count up)

• Tracing # of view counts – for every hour– 2010.08.15 ~ 2010.08.26– for talkative users

Views of MLB (News)

days

Views of O_CONNECTION (Humor)

days

views

Views of ladygaga (Celebrity)

days

Spreading Tree Analysis

• Using a connected tree from source user• Remove loops, multiple edges

Spreading tree reconstruction

• “RT @Somebody : blah blah”

• General messages

• Reply

Information Spreading Pattern

The median value of properties for trees

Cascade size 17

Max. depth 3

Median depth 1.5

Width 10

Single-edge frac-tion

0.125

Source contribu-tion

0.4375Diffusion trees in Twitter are wide and shal-low.

The source plays an important role in infor-mation diffusion

Source vs the Others

Same # of Tweets, Different Pattenrs of Diffusion

Response Probability

27

Social capacity of human beings

• Dunbar’s number

28

Dunbar’s number

Behavioral and brain scineces, 16(4):681–735, 1993

The maximum number of social relations managed by modern human is 150.

29

#(friends) stimulate interaction?

The more friends one has (up to 200), the more active one is.Median

#(sent msgs)

30

Twitter activity vs # of followings

31

Caveats

• Not complete from an ego-centric perspective

32

Break-up

33

Two Sides of Relationship

• Formation and Dissolution– Formation tradiationally well studied– Dissolution hardly much

• Why?– Hard to obtain data

• Proxy for dissolution– No exchange of email [Kossinet09]

34

Two Questions We Raise

• How prevalent is unfollow?

• Why do people unfollow?

35

Four Types of Tweet

• TweetPSSM is now starting!

• Reply@Virgilio Fantastic Workshop! Thanks for having me!

• MentionI am attending PSSM organized by @Virgilio and @PK!

• RetweetAt UFMG till tomorrow! RT @Virgilio PSSM is now start-ing!

Proportion of Tweet Types

Users become more informational than interactive as the number of followees increases

How Prevalent Is Unfollow?

37

Follows and Unfollows

Unfollow is prevalent!

39

Unfollow frequent

• Mostly singular– 66% of unfollows are the only unfollow of the day

• But often clustered– 10% with 5 or more other unfollows

• On average– 90% of time intervals between days of unfollow is

less than 9 days

40

Communication partner

• Reciprocal and interactive users– Exchange of a mention, a reply, or a retweet and

vice versa

#Comm Partners vs. #Followees

42

Passive Nature of Follow

• 85.6% relationships involve no activity• 96.3% involve 3 or fwer• Who unfollows?– Remove 85.6% of no activity and among those

with any activity unfollowed relationships involves less activity than unbroken relationships

Unfollow ratio vs. ego-centric ordering of re-lationship establishments

# Followees vs. # Unfollowees

More Retweets/Favorites Less Likely to Be Unfollowed

The overlap of relationships vs. unfollow ra-tio

Why Do People Unfollow?

47

48

Interviews

Q1: Why a participant decided to unfollow.Q2: Whether s/he thought the unfollowee was aware of being unfollowed.Q3: If s/he broke off on other OSNs. Difference?Q4: If s/he followed corporate accounts.Q5: Choose 10 users s/he would never unfollow

Demographics of 22 interviewees

50

Q1: Motivations behind Unfollow

• Burst (39)– Burst-only (13), Unintersting topic (10), Mundane

details (6), Automatically generated (4), Conversa-tion (2), Politics (2), Different Views (1), Complains (1)

51

Q2: Awareness of Being Unfollowed

• A half of respondents stated that they thought unfollowees were aware of being unfollowed.– They did not know unfollowees in person– They got used to unfollowing– Unfollow was easy

• The other half– Unfollowees had too many followers to notice– No convenient interface to track it– They did not track themselves

52

Q3: Break-up on other OSNs?

• Not common

53

Q4: Corporate Accounts

• 8 out of 22 follow corporate accounts– 5 kept following– Motivaiton = expectation of prize winning– They didn’t mind occasional ad tweets, but unfol-

low if ads come in bursts– Some only participate if all participants received a

gift

54

Q5: Whom Not to Unfollow?

• Most respondents chose intimate friends• Some chose their role models

55

Conclusions

• Just a tip of an iceberg for computational – social science– journalism– political science– archeology– literature study– linguistics