1. Mul$-­‐level
Analysis
on
Structures
and
Dynamics
of
OSN
Haewoon
Kwak
Department
of
Computer
Science,
KAIST
Ph.
D.
thesis
defense
April
12th
2011
Advisor:
Sue
Moon
1

2. Outline
• Background:
data
sources
for
social
research
– Tradi$onal
methodologies
developed
in
sociology
– Electronic
footprints
• Complex
structures
and
dynamics
of
OSN
• Mul$-­‐level
approach
Individual
level
Dyad
level
Community
level
Network-­‐wide
level
• Summary
&
future
direc$on
2

3. • Surveys
• Ques$onnaires
• Archives
• Observa$ons
• Experiments
• Issues
of
data
scalability,
quality,
&
measurement
[Marsden
90]
Methodologies
developed
in
sociology
3

4. Self-­‐report
vs.
observed
Eagle
N,
Pentland
A,
LazerD
(2009)
Inferring
friendship
network
structure
using
mobile
phone
data.
PNAS,
106:15274–15278. 4

5. The
emergence
of
electronic
footprints
Tracking
People's
Electronic
Footprints,
Science
10
November
2006:
Vol.
314
no.
5801
pp.
914-­‐916
5

6. • Mobile
phone
call
logs
[J.
P.
Onnela
et
al.
07]
• E-­‐mail
logs
[Kossinets
&
D.
Waas
09]
• Online
‘friend’
rela$onships
[Ahn
et
al.
07]
• Facebook
wall
pos$ngs
[S.
Golder
et
al.
07]
• Conversa$ons
on
MSN
[Leskovec
&
Horvitz
07]
• Photos
in
Flickr
[Marlow
et
al.
07]
Forms
of
electronic
footprints
6

7. The
era
of
online
social
networks
7

8. • Facebook
– 30B
pieces
of
new
informa$on
/
month
• Twiaer
– 1B
messages
/
week
– 7TB
/
day
– “300GB
while
I
give
this
talk”
Tremendous
volume
of
records
hap://www.facebook.com/press/info.php?sta$s$cs
NoSQL
at
Twiaer
(NoSQL
EU
2010) 8

9. • Establish
online
‘friend’
rela$onship
• Send
a
message
• Send
a
gil
• Upload
a
photo
• Share
one’s
loca$on
• Play
a
game
Rich
behavior
in
OSN
9

10. • New
source
that
reveals
human
nature
– As
a
miniature
of
human
society
– e.g.
verifying
balanced
theory
&
weak
$e
hypothesis
• Virtual
world
interac$ng
with
a
real
world
– Elec$on
campaign
Q:
Why
we
study
OSN?
A:
OSN
is
…
10

11. • New
source
that
reveals
human
nature
– As
a
miniature
of
human
society
– e.g.
verifying
balanced
theory
&
weak
$e
hypothesis
• Virtual
world
interac$ng
with
a
real
world
– Elec$on
campaign
– Gossip
propaga$on
Q:
Why
we
study
OSN?
A:
OSN
is
…
11

12. • New
source
that
reveals
human
nature
– As
a
miniature
of
human
society
– e.g.
verifying
balanced
theory
&
weak
$e
hypothesis
• Virtual
world
interac$ng
with
a
real
world
– Elec$on
campaign
– Gossip
propaga$on
– Reputa$on
management
Q:
Why
we
study
OSN?
A:
OSN
is
…
12

13. • New
source
that
reveals
human
nature
– As
a
miniature
of
human
society
– e.g.
verifying
balanced
theory
&
weak
$e
hypothesis
• Virtual
world
interac$ng
with
a
real
world
– Elec$on
campaign
– Gossip
propaga$on
– Reputa$on
management
– Money
exchange
Q:
Why
we
study
OSN?
A:
OSN
is
…
13

14. Complex
structures
and
dynamics
(1)
American
Journal
of
Sociology,
Vol.
100,
No.
1."Chains
of
affec$on:
The
structure
of
adolescent
roman$c
and
sexu
al
networks
“
(2004)
Facebook
social
graph
by
Facebook
Data
team
14

15. Complex
structures
and
dynamics
(2)
Annual
Review
of
Sociology,
Vol.
30
“THE
‘NEW’
SCIENCE
OF
NETWORKS”
(2004)
Nature
reviews
gene$cs,
Vol.
5,
“Network
biology:
under
standing
the
cell's
func$onal
organiza$on”
(2004)
WWW,
“Analysis
of
Topological
Characteris$cs
of
Huge
Online
Social
Networking
Services”
(2007)
15

16. • OSN’s
evolu$on…
only
#(users)?
Complex
structures
and
dynamics
(3)
16

17. • Mul$-­‐level
approach
to
OSN
• Mul$ple
views
focusing
on
different
en$$es
Individual
Dyad
Community
Network-­‐wide
Our
approach
–
Divide
and
conquer
17

18. • Each
level
has
its
own
inherent
resolu$on
– to
capture
elemental
processes
– to
understand
the
complex
structures
and
dynamics
as
a
combina$on
of
findings
across
mul$ple
levels
Strong
points
of
mul$-­‐level
approach
18

19. e.g.
How
a
user
has
posi$onal
power?
A
story
begins
with
an
individual
19

20. An
individual
has
proper$es
20

21. Individuals
have
proper$es
21

22. Similar
proper$es
drive
rela$onships
22

23. Densely
connected
individuals
form
a
community
23

24. Individuals
interact
with
others
in
the
same
community
24

25. Some
individuals
connect
to
many
people
in
a
comm.
25

26. Some
individuals
bridge
different
communi$es
26

27. • OSN’s
evolu$on…
only
#(users)?
Complex
structures
and
dynamics
(3)
27

28. • Different
perspec$ves
of
the
evolu$on
of
OSN
– Increasing
avg.
Mme
on
sites
– Increasing
#
of
friends
– Diversifying
types
of
rela$onship
– Increasing
#
of
cohesive
groups
– Increasing
density
of
the
network
– Shortening
the
avg.
diameter
of
the
network
– Absorbing
other
networks
Complex
structures
and
dynamics
(3)
Complex
structures
and
dynamics
(3)Complex
phenomena
in
mul$-­‐level
28

29. • Cannot
be
answered
in
one-­‐level
only
• Thus,
we
tackle
it
step
by
step
from
microscopic
to
macroscopic
view
– Individual
level
– Dyad
level
– Community
level
– Network-­‐wide
level
29
Complex
structures
and
dynamics
of
OSN

30. • Mul$-­‐level
analysis
of
structures
and
dynamics
– Personal
preferences
and
friend
recommenda$on
– Rela$onship
dynamics
– Consistent
community
iden$fica$on
– Interplay
between
structures
and
dynamics
Overview
of
this
thesis
30

31. Overview
of
this
thesis
31
2008
2009 2010 2011
Comparison
of
Online
Social
Rela7ons
In
Terms
of
Volume
vs.
Interac7on:
A
Case
Study
of
Cyworld
Hyunwoo
Chun,
Haewoon
Kwak,
Young-­‐Ho
Eom,
Yong-­‐Yeol
Ahn,
Sue
Moon,
and
Ha
woong
Jeong,
The
8th
ACM
SIGCOMM
Conference
on
Internet
Measurement
(IMC).
2008.

32. Overview
of
this
thesis
32
2008 2009
2010 2011
Connec7ng
Users
with
Similar
Interests
Across
Mul7ple
Web
Services
Haewoon
Kwak,
Hwa-­‐Yong
Shin,
Jong-­‐Il
Yoon,
and
Sue
Moon,
The
3rd
Interna$onal
AAAI
Conference
on
Weblogs
and
Social
Media
(ICWSM),
Poster,
2009.
Mining
communi7es
in
networks:
A
solu7on
for
consistency
and
its
evalua7on
Haewoon
Kwak,
Yoonchan
Choi,
Young-­‐Ho
Eom,
Hawoong
Jeong,
and
Sue
Moon.
The
9th
ACM
SIGCOMM
Conference
on
Internet
Measurement
(IMC),
2009.

33. Overview
of
this
thesis
33
2008 2009 2010
2011
What
is
TwiKer,
a
Social
Network
or
News
Media?
Haewoon
Kwak,
Changhyun
Lee,
Hosung
Park,
and
Sue
Moon,
The
19th
interna$onal
conference
on
World
wide
web
(WWW),
2010
Finding
Influen7als
Based
on
the
Temporal
Order
of
Info.
Adop7on
in
TwiKer
Changhyun
Lee,
Haewoon
Kwak,
Hosung
Park,
and
Sue
Moon,
The
19th
interna$onal
conference
on
World
wide
web
(WWW),
Poster,
2010
Ph.
D.
Thesis
Proposal

34. Overview
of
this
thesis
34
2008 2009 2010 2011
Fragile
Online
Rela7onship:
A
First
Look
At
Unfollow
Dynamics
In
TwiKer
Haewoon
Kwak,
Hyunwoo
Chun,
and
Sue
Moon,
The
29th
interna$onal
conference
o
n
Human
factors
in
compu$ng
systems
(CHI),
2011.

35. Individual
&
dyad-­‐level
view:
Personal
preferences
&
friend
recommenda$on
Connec7ng
Users
with
Similar
Interests
Across
Mul7ple
Web
Services
Haewoon
Kwak,
Hwa-­‐Yong
Shin,
Jong-­‐Il
Yoon,
and
Sue
Moon,
The
3rd
Interna$onal
AAAI
Conference
on
Weblogs
and
Social
Media
(ICWSM),
Poster,
2009.
다종 웹 서버 간 유사 사용자 추출 시스템 및 그 방법
Sue
Moon,
Haewoon
Kwak,
Hwa-­‐Yong
Shin,
and
Jong-­‐Il
Yoon,
Korean
Patent
No.
10-­‐1010997,
2011.
35

36. • Groups
in
OSN
– Mo$vate
user
ac$vi$es
– Share
common
interests
and
offline
commonali$es
Mo$va$on
Blackbox
But,
restricted
within
the
boundary
of
each
service
36

37. • To
recommend
people
who
share
interests
across
various
kinds
of
web
services
• To
support
many
web
services
without
modifica$on
• Not
to
burden
users
with
addi$onal
profile
management
Our
goals
37

38. • Inferring
interests
from
tags
– Free
keywords
to
describe
user-­‐generated
contents
– Publicly
accessible
– Simple
format
(plain
text)
Key
insights
38

39. Tags
are
supported
by
many
services
39

40. • Tags
from
6
various
services
• Refinement
by
WordNet
API
Datasets
Contents Uniq.
users Uniq.
tags Avg.
tags
Del.icio.us Bookmark 40,072 1,092,534 227.2
Flickr Photo 6,366 71,724 32.4
YouTube Video 9,481 171,990 56.5
LiveJournal Blog 49,792 729,975 44.49
Last.FM Music 54,464 95,901 10.95
AllBlog Blog 24,559 383,374 44.04
40

41. • Highly
skewed
popularity
– Top
20%
of
tags
associated
with
90%
of
items
• Service-­‐dependent
– About
20%
of
tags
belong
to
more
than
one
service
• Frequently
change
over
$me
Findings
about
user
interests
41

42. • Tag
weight
assignment
– Absolute
number
of
$mes
a
tag
is
used
– Normalized
number,
N,
of
$mes
a
tag
is
used
– N-­‐idf
• Similarity
calcula$on
between
two
tag
sets
– Sum
of
the
weights
of
common
tags
– Cosine
similarity
based
on
the
vector
model
Recommenda$on
algorithms
42

43. User
study
of
algorithm
evalua$on
N-­‐idf
+
cosine
similarity
works
43

44. • Flickr
and
YouTube
users
find
good
matches
in
other
services
• LiveJournal
aaracts
users
from
other
services
beaer
than
any
other
service
Condi$onal
prob.
of
recommenda$on
44

45. Dyad-­‐level
view:
Social
rela$onship
dynamics
Comparison
of
Online
Social
Rela7ons
In
Terms
of
Volume
vs.
Interac7on:
A
Case
Study
of
Cyworld
Hyunwoo
Chun,
Haewoon
Kwak,
Young-­‐Ho
Eom,
Yong-­‐Yeol
Ahn,
Sue
Moon,
and
Hawoong
Jeong,
The
8th
ACM
SIGCOMM
Conference
on
Internet
Measurement
(IMC’08).
2008.
Fragile
Online
Rela7onship:
A
First
Look
At
Unfollow
Dynamics
In
TwiKer
Haewoon
Kwak,
Hyunwoo
Chun,
and
Sue
Moon,
The
29th
interna$onal
conference
on
Human
factors
in
compu$ng
systems
(CHI’11),
2011.
45

46. • Exchange
of
guestbook
messages
in
Cyworld
• Rela$onship
dissolu$on
by
unfollow
in
Twiaer
Rela$onship
dynamics
observed
from
46

47. • Most
popular
online
SNS
in
Korea
(22M
users)
• Guestbook
is
the
most
popular
feature
Part
I:
guestbook
log
analysis
in
Cyworld
47

48. • Online
‘friends’
rela$onship
– Needs
no
more
cost
once
established
– Commonly
mutual
– All
online
friends
are
considered
equally
Mo$va$on
Thus,
online
‘friends’
are
not
enough
to
represent
soci
al
rela$onships
at
that
$me
48

49. • Direc7onality
and
strength
of
user
interac$ons
reveal
more
meaningful
rela$onships
than
online
‘friends'
do
Key
insights
49

50. From
logs
to
the
ac$vity
network
<
From,
To,
When
>
<A,
C,
20040103T1103>
<B,
C,
20040103T1106>
<C,
B,
20040104T1201>
<B,
C,
20040104T0159>
CA
B
1
2 1
Directed
&
weighted
“AcMvity
network”
8
billion
messages
Graph
construc$on
50

51. • Have
power-­‐law
degree
distribu$on
– A
few
number
of
high-­‐degree
nodes
– A
large
number
of
low-­‐degree
nodes
• Have
common
characteris$cs
– Short
diameter
– Fault
tolerant
Most
social
networks
Nature
Reviews
GeneMcs
5,
101-­‐113,
2004
51

52. Mul$-­‐scaling
behavior
implies
heterogeneous
rela$ons
52

53. Weighted
clustering
coefficient
PNAS,
101(11):3747–3752,
2004
i1 w
=
10
w
=
1
i2
48
5.6
)
2
)11()110(
(
)13(12
1
1 =
+++
−
=w
iC
48
11
)
2
)110()101(
(
)13(12
1
2 =
+++
−
=w
iC
w
i
w
i CC 21 <
53

54. • In
ac$vity
network
Cw=0.0965
<
C=0.1665
Weighted
clustering
coefficient
Edges
with
large
weights
are
less
likely
to
form
a
triad
i1 i2
54

55. Degree
correla$on
of
social
network
degree
avg.
degree
of
neighbors
M.E.J.
Newman.
AssortaMve
mixing
in
networks,
Phys.
Rev.
Le`.
89,
208701
(2002).
“Assorta$ve
mixing”
55

56. Degree
correla$on
of
ac$vity
network
Assorta$ve
mixing
is
observed
56

57. Reciprocity
in
user
ac$vi$es
y=x
-­‐
Highly
reciprocal
-­‐
Quan$ta$ve
proof
of
spammers
57

58. • Do
users
interact
evenly
with
all
friends?
Disparity
Journal
of
Physics
A:
MathemaMcal
and
General,
20:5273–5288,
1987.
For
node
i,
Y(k)
is
average
over
all
nodes
of
degree
k
58

59. Interpreta$on
of
Y(k)
Nature
427,
839
–
843,
2004
Communicate
evenly Have
dominant
partner
59

60. Disparity
in
user
ac$vi$es
Communica$on
paaern
changes
by
#(partners)
60

61. • 13
possible
interac$on
paaerns
with
3
users
• Propor$ons
of
each
paaern
(mo$f)
determine
the
characteris$c
of
the
en$re
network
Network
Mo$fs
Network
MoMfs:
Simple
Building
Blocks
of
Complex
Networks,
Science,
298(5594):824-­‐827,
2002 61

62. Mo$f
analysis
in
complex
networks
Superfamilies
of
Evolved
and
Designed
Networks,
Science, 303(5663):1538-1542, 2004
In
social
networks,
triads
are
likely
to
be
observed
62

63. Network
mo$fs
in
user
ac$vi$es
Triads
are
common
in
Cyworld
63

64. Network
mo$fs
in
user
ac$vi$es
Not
in
social
networks,
but
in
OSN
64

65. Dunbar’s
number
Behavioral
and
brain
sciences,
16(4):681–735,
1993
The
maximum
number
of
social
rela$ons
managed
by
modern
human
is
150
65

66. #(friends)
s$mulate
interac$on?
The
more
friends
one
has
(up
to
200),
the
more
ac$ve
one
is
Median
#(sent
msgs)
66

67. Time
interval
between
messages
Nature,
435:207–211,
2005
Proceedings
of
WWW,
2008
intra-­‐session
inter-­‐session
daily-­‐peak
67

68. Part
II:
unfollow
analysis
in
Twiaer
“Stop
following”
68

69. • Rela$onship
forma$on
and
dissolu$on
– Forma$on
has
received
much
aaen$on
– Dissolu$on
hardly
much
due
to
the
lack
of
data
• Proxy
such
as
a
disappearance
of
communica$on
– is
difficult
to
capture
all
communica$on
means
– regards
the
absence
of
an
event
as
strictly
inten$onal.
Mo$va$on
69

70. • Unfollow
in
Twiaer
is
an
explicit
expression
of
rela$onship
dissolu$on
Key
insights
70
&
research
ques$ons
• What
are
the
characteris$cs
of
unfollow?
• Why
do
people
unfollow?

71. • 1.2M
Korean-­‐speaking
users
detected
by
– Korean
in
tweets,
bio,
loca$on,
or
screen
name
• Daily
snapshots
of
follow
rela$onships
– G(I):
June
25th
to
July
15th,
2010
– G(II):
August
2nd
to
August
31st,
2010
Datasets
71

72. • Increasing
#
of
users
– G(I):
870,057
+7,599/day
– G(II):
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
Growing
Korean
social
graphs
72

73. Rela$onship
forma$on
and
dissolu$on
73

74. • Reciprocity
of
the
rela$onships
• Dura$on
of
a
rela$onship
• Followees’
informa$veness
• Tie
strength
4
Factors
to
unfollow
74

75. Non-­‐reciprocal
links
are
more
fragile`
P(broken)
=
0.1228
P(broken)
=
0.0529
P(broken)
=
0.2345
75
<
<

76. Newer
links
are
more
fragile
76

77. Non-­‐informa$ve
links
are
more
fragile
Random
links
77

78. Weak
links
are
more
fragile
78

79. • 85.6%
of
links
do
not
involve
any
single
reply,
men$on,
or
retweet
– 96.3%
involve
3
or
fewer
• People
just
subscribe
others’
tweets
passively
79
Volume
of
ac$vity
is
not
a
good
proxy

80. Demographic
of
22
par$cipants
11
Male,
11
Female
80

81. • Burst
tweets
• Uninteres$ng
topics
• Mundane
details
of
daily
life
• Poli$cs
Interviews
about
mo$va$on
81

82. • Burst
tweets
are
likely
to
lead
unfollow
82
Confirmed
by
data
Pearson
corr.
=
0.0554 Pearson
corr.
=
0.5833
All
users
Followee
<
200
Go
to
summary

83. Community-­‐level
view:
Consistent
community
iden$fica$on
Mining
communi7es
in
networks:
A
solu7on
for
consistency
and
its
evalua7on
Haewoon
Kwak,
Yoonchan
Choi,
Young-­‐Ho
Eom,
Hawoong
Jeong,
and
Sue
Moon.
The
9th
ACM
SIGCOMM
Conference
on
Internet
Measurement
(IMC’09),
2009.
Consistent
Community
Iden7fica7on
in
Complex
Networks
Haewoon
Kwak,
Young-­‐Ho
Eom,
Yoonchan
Choi,
Hawoong
Jeong,
and
Sue
Moon,
Preprint
(arXiv:0910.1508v2),
2009
83

84. Which
par$$oning
is
beaer?
84

85. • eii:
ra$o
of
the
number
of
links
between
nodes
belonging
to
community
i
over
all
links
• ai:
ra$o
of
ends
of
edges
that
are
aaached
to
ver$ces
in
community
i
Modularity,
Q
85

86. High
Q
is
beaer
Q=
0.41979Q=
0.380671
<
86

87. • Greedy
algorithms
to
maximize
Q
– Widely
used
for
a
few
tens
of
million
nodes
network
– By
nature,
no
guarantee
to
find
global
maximum
– Local
maxima
=
Inconsistent
par$$oning
Mo$va$on
87

88. • Measure
the
level
of
inconsistency
• Develop
a
new
method
to
achieve
consistency
Our
goals
88

89. Datasets
of
12
networks
89

90. • The
likelihood
of
a
pair
of
nodes
resul$ng
in
the
same
community
over
N
runs
Measure
1:
Pairwise
membership
prob.
90

91. 91
Example
of
pairwise
membership
prob.

92. 92
Example
of
pairwise
membership
prob.
p.m.p.
=
2/2
p.m.p.
=
1/2

93. • To
quan$fy
the
network-­‐wide
consistency
Measure
2:
Consistency,
C
93

94. Consistency
in
12
networks
94

95. • Every
edge
has
pairwise
membership
prob.
• High
pairwise
membership
prob.
indicates
that
two
nodes
are
likely
to
be
in
the
same
community
• Weighted
version
of
exis$ng
algorithms
place
edges
of
high
weight
within
the
community
Intui$ons
behind
new
algorithm
95

96. 1. Aler
a
cycle
of
N
runs,
– Calculate
pairwise
membership
prob.
of
each
edge
– Assign
pairwise
membership
prob.
to
edge
weight
2. Return
to
another
cycle
of
N
runs
with
an
weighted
network
3. Go
to
1.
again
un$l
C
>=
Ƭ
(predefined
threshold)
Our
new
algorithm
96

97. Convergence
of
C
97

98. Network-­‐wide
view:
Interplay
between
structures
and
dynamics
What
is
TwiKer,
a
Social
Network
or
News
Media?
Haewoon
Kwak,
Changhyun
Lee,
Hosung
Park,
and
Sue
Moon,
The
19th
interna$onal
conference
on
World
wide
web
(WWW),
2010
Finding
Influen7als
Based
on
the
Temporal
Order
of
Info.
Adop7on
in
TwiKer
Changhyun
Lee,
Haewoon
Kwak,
Hosung
Park,
and
Sue
Moon,
The
19th
interna$onal
conference
on
World
wide
web
(WWW),
Poster,
2010
98

99. Mo$va$on
In
Twiaer
“I
follow
you”
In
most
OSNs
“We
are
friends”
99

100. • Measure
how
one-­‐way
rela$onship
affects
the
network-­‐wide
structures
and
dynamics
Our
goals
100

101. • 41.7M
user
profiles
• 1.47B
follow
rela$onships
• 4,262
trending
topics
• 106M
tweets
men$oning
trending
topics
– Spam
tweets
removed
by
CleanTweets
Datasets
101

102. • Only
22.1%
of
user
pairs
follow
each
other
• Much
lower
than
– 68%
on
Flickr
– 84%
on
Yahoo!
360
– 77%
on
Cyworld
guestbook
Low
reciprocity
in
follow
rela$onships
102

103. Plenty
of
super
hubs
103

104. Degree of Separation
104
The
avg.
path
length
=
4.1

105. Timely
trending
topics
54.3%
“headline
news”
31.5%
“ephemeral”
105

106. 106
Empirical
retweet
trees

107. Boos$ng
audience
by
retweets
107

108. 35%
of
RTs
<
10
min.,
55%
<
1
hr
108

109. 109
Powers
from
structures
and
dynamics

110. • Mul$-­‐level
analysis
on
structures
and
dynamics
– Individual-­‐level
– Dyad-­‐level
– Community-­‐level
– Network-­‐wide
level
Summary
110

111. • Elemental
processes
in
each
level
– Personal
preferences
and
friend
recommenda$on
– Rela$onship
dynamics
– Consistent
community
iden$fica$on
– Interplay
between
structures
and
dynamics
Ques$ons
we
raise
111

112. • From
new
algorithms
to
analyses
– New
algorithms
and
its
evalua$on
for
recommending
those
who
have
similar
interests
across
services
– Analysis
of
actual
interac$ons
in
contrast
to
friends
rela$onships
in
a
huge
OSN,
Cyworld
– Quan$ta$ve
and
qualita$ve
analysis
of
unfollow
– New
metrics
for
evalua$on
of
consistency
in
communi$es
iden$fied
by
exis$ng
algorithms
– Analysis
of
structures
and
dynamics
of
a
huge
directed
network,
Twiaer
Our
contribu$ons
112

113. • The
interplay
among
the
parallel
social
networks
from
mul$ple
services
• Conflic$ng
informa$on
pathways
in
social
media
Future
direc$ons
113
MulMrelaMonal
organizaMon
of
large-­‐scale
social
networks
in
an
online
world,
PNAS
107(31):13636-­‐13641

114. • We
aggregate
parallel
social
networks
captured
in
each
service
and
construct
a
big
network.
• We
observe
the
differences
between
parallel
networks
• We
observe
the
interplay
between
parallel
networks
such
as
node
migra$on
• We
compare
user’s
posi$onal
power
between
parallel
networks
Our
goals
114

115. • Some
people
selec$vely
propagate
informa$on
by
their
preferences
• Different
transmission
path
– Some
know
correct
info.,
but
others
not
Conflic$ng
informa$on
pathways
115

116. Thank
you
116