Carnegie Mellon School of Computer Science Modeling Website Popularity Competition in the...

Carnegie MellonSchool of Computer Science

Modeling Website Popularity Competition in the Attention-Activity

Marketplace

Bruno Ribeiro Christos FaloutsosCarnegie Mellon University

WSDM 2015February 5, 2015

Bruno Ribeiro ([email protected])Carnegie MellonSchool of Computer Science

2

Motivation

Bought for $580 million

in 2005

Sold for $35 million in

2011

*source: TechCrunch

MySpace’s demise

Sold

Hi5Friendster

Summer 2008

Summer 2008


3

Predicting Network Popularity w/

Competition

Goal:

Why:

??

?

Summer 2008

Fract

ion o

f A

ctiv

e U

sers

(t)

-$545 million dollars


4

new adopters/semester

MySpace.com until late 2008

No Structural Telltale Signs of MySpace’s Demise

Matches known theoretical behaviorMansfield’61, Rogers’03, Bass’69

Total Adopters

0Innovators2.5 %

EarlyAdopters13.5 %

EarlyMajority34 %

LateMajority34 %

Laggards16 %

And no abrupt topological changes in 2008

MySpace graph until recently:• 35 million users

★ http://www.myspace.com


5

Economics:◦ (Mansfield ’63)◦ (Katz&Shapiro’85) ◦ (Farrell&Saloner

’86)◦ (Choi ’94)◦ (Arthur ’94)

Marketing:◦ (Bass ’69)◦ (Fisher&Pry ’71)

Fract

ion o

f A

ctiv

e U

sers

(t)

Summer 2008

Background: Vast Adoption Literature

Computer Science: (Kempe et al ’03) (Zhao et al., IMC’12) (Leskovec et al.,

SIGKDD’08) (Ugander et al., PNAS’12) (Aral&Walker,

Science’12)

Sociology:o (Ryan&Gross’49)o (Everett ’62, ’03)o (Rogers ’03)o (Centola ’12)

Adopt ≠ Active

We know how to model adoption

But how to model attention?


6

1969 - H. A. Simon on information overload:

Information consumes attention (time)

◦ Information-rich world Attention-poor world

Systems that “talk” more than “think” exacerbate

information overload

Attention & Information Overload


Facebookattention& activity

7

How Facebook Talks (a lot)


8

My Attentio

n

My Content

Friends’ Attentio

n

Friends’ Content

Positive & Negative Attention Loops

Positive

Growth

Negative Growth

WebsiteSurvives

WebsiteDies


9

Proposed Competition Model


10

Proposed State Space

states ∅yUy Ay Iy

∅x

UxAxIx

Both x & y Only y for now Only x for now

Neither for now Not interested in x/y

User Subscribes to:

WebsiteKillers

Never! Unaware Active Inactive


11

Model Transitions

User Competition


12

Isolated Network Transitions

states ∅yUy Ay Iy SUM

∅x

Ux S(U, )∅ (t)

Ax S(A, )∅ (t) S(A,*)(t)

Ix S(I, )∅ (t)

Media/Marketing gets new subscribers Word of mouth gets subscribers

Network x activity attracts user back

Other activitytakes user away from x

Details


13

Model Predictions with Constant Distraction Factor

Predicting Online Social

Network Popularity

◦ How websites do on

their own

◦ Evolution with

competition

Outline


14

Signatures of Self-Sustainability

Long-term User Activity[Ribeiro’14]

Relative attractiveness of activity

Fract

ion

of

Act

ive U

sers

(t)

Fract

ion

of

Act

ive U

sers

(t)

t t

≅1MySpace


15

Signatures of Popularity Growth

Model prediction: Signatures of activity growth[Ribeiro’14]

t tFract

ion

of

Act

ive U

sers

(t)

Fract

ion

of

Act

ive U

sers

(t)

Wor

d of

mou

th

Med

ia &

Mar

ketin

g


16

Predictions with Changing Distractions

t

Predicting Online Social

Network Popularity

◦ How websites do on

their own

◦ Evolution with

competition

Outline


17

Two observable states: S(A,*)(t) and S(*,A)(t)

Remaining states latent & treated as parameters

Data from Alexa.com

Fitted with Levenberg–Marquardt algorithm using squared error

Model Fitting

S(A,*)(t)

S(I,*)(t)?S(U,*)(t)?S(A,I)(t)?S(A,A)(t)?S(A,U)(t)?S(A, )∅ (t)?S(I, )∅ (t)?S(U, )∅ (t)?

friendster.com

Fract

ion

of

Act

ive U

sers

(t)

Details


18

MySpace

Friendster

Multiply

Hi5

Facebook Introduces Wall

Distraction of Concurrent UsersIncreases as Facebook Introduces

Wall


19

Results (Facebook x MySpace)

t

value


20

Results (Facebook x Multiply)

t

value

Multiply


21

Results (Facebook x Hi5)

t

value

Hi5


22

Results (Facebook x Hi5)

Model Captures Coexistence and Then

Death of Facebook Competitors

Too Many Concurrent Users = Fragility


23

Attention feedback helps predict /

understand network survival

No marketing can save from negative

attention loop

Insights from model

◦ E.g.: Metrics of node centrality should take

attention feedback into account

Conclusions


24

Thank you!

Questions?

http://www.cs.cmu.edu/~ribeiro@brunofmr

Carnegie Mellon School of Computer Science Modeling Website Popularity Competition in the...

Documents

Transcript of Carnegie Mellon School of Computer Science Modeling Website Popularity Competition in the...