1

Fragile Online Relationship:A First Look at Unfollow Dynamics in Twitter

Haewoon Kwak Hyunwoo Chun Sue MoonDepartment of Computer Science, KAIST

The 29th international conference on Human Factors in Computing Systems (CHI)

10th May 2011

2

• Two basic processes in social network evolution

• But we don’t talk about breakup

Building and breaking up a relationship

??

3

• Many records of relationship establishment– Sending and accepting a friend request– Writing a paper together for the first time– Making a call for the first time

• Breaking up a relationship– Hard to define and capture

Lack of relationship breakup data

4

• Disappearance of e-mail exchange

A proxy for relationship breakup

But they were doing instant messaging.

“Alice & Bob had exchanged e-mails frequently. At some point they didn’t do any more.”

A proxy is not always accurate!

5

• Intentional action to break a relationship• No need for an approval• No notification to the unfollowed

Unfollow in Twitter

6

• RQ1: What are the characteristics of unfollow?• RQ2: Why do people unfollow?

Research questions

7

• Not feasible to crawl the entire Twittersphere• Collecting 1.2M Korean-speaking users identified

by Korean in tweets, bio, location, or name

Identify a group with common culture

8

• Collect daily snapshots of follow networks– G(I): June 25th to July 15th, 2010– G(II): August 2nd to August 31st, 2010

• Compare two consecutive follow networks and detect removed relationships

Obtaining unfollow records

June 25th June 26th

…Removed relationship

9

• Increasing # of users– G(I): 718,077 → 870,057 +7,599/day– G(II): 956,261 → 1,203,196 +8,515/day

• Increasing (high) reciprocity– G(I): 56 → 58%– G(II): 61 → 62%

• Increasing avg. # of followees– 59.7 → 75.7

Korean follow network grew fast

10

• RQ1: What are the characteristics of unfollow?• RQ2: Why do people unfollow?

Research questions

11

• 43% of active users unfollow at least once during 51 days

• Average number of unfollows per person– 15.4 in G(I)– 16.1 in G(II)

People unfollow frequently

12

• Reciprocity of relationships• Duration of a relationship• Followee’s informativeness• Overlap of relationships

Factors that correlate w/ unfollow

13

One-way relationships are fragile

P(broken) = 0.1228

P(broken) = 0.0529

P(remaining will be broken|one is broken) = 0.2345

<<

One-way

Reciprocal“Emotional closeness”

14

Newer relationships are more fragile

Is a user less likely to unfollow those whom the user has been following for some time?

15

Informativeness of users

16

Non-informative relationships are fragile

17

Strong ties & weak ties

http://bokardo.com/archives/weak-ties-and-diversity-in-social-networks/

18

Weak ties are fragile

19

• 85.6% of relationships do not involve any single reply, mention, or retweet– 96.3% involve 3 or fewer

• People just subscribe to others’ tweets passively

No interaction ≠ breakup

20

• RQ1: What are the characteristics of unfollow?• RQ2: Why do people unfollow?

Research questions

21

22 online & face-to-face interviews

11 Male, 11 Female

• Recruited by word-of-mouth• Semi-structured• Logging & camera recording

22

1. Burst tweets2. Tweets about uninteresting topics3. Tweets about mundane details of daily life

• Automatically generated tweets (4sq)

4. Tweets about political issues

Top reasons in unfollow

23

Application: Practical usesBurst tweets, automatically generated tweets, …

24

• Existing models for growing networks deal with new relationship formation only– Preferential attachment model– Forest fire network model

• We add another dimension to network evolution models from the perspective of relationship breakup

Application: Theoretical models

25

• Generalization of this work– One user group study

• Online relationship not equal to social friendship

Discussion

26

• Unfollow is a clear sign of breaking a relationship– None of interaction is not a good indicator of breaking

a relationship due to user’s passivity in Twitter

• The first study of the unfollow behavior– Quantitative data analysis revealed its characteristics– Qualitative interviews found its motivation

Summary

27

28

Backup slides

29

How to know the reciprocity of follows

30

• Burst tweets are likely to lead unfollow

Interviews confirmed by data

Pearson corr. = 0.5833

Followee < 200

31

32

33

34

35