nyu stern spring 2013 - mitweb.mit.edu/sinana/www/digital analytics & strategy 2013 -...

NYU Stern Spring 2013

Digital Analytics & Strategy


Sessions 3 & 4: Network AnalyticsNetworks, Influence and Social Media Analytics

Prof. Sinan Aral

Digital Analytics & Strategy: Sessions 3 & 4

Learning Objectives: Network Analytics

1. Understand why (Social) Networks are so critical to demand prediction and marketing.

2. Understand Economic Network Effects and why they are

so essential to Digital Strategy.

3. Understand the importance of Causal Statistical

Estimation in effective Social Network Marketing efforts.

4. Understand Viral Product Design and its implications for

a) Social Contagion in Product Adoption, b) Sustained

Product Use and c) the relationship between the two.

5. Consider how to identify influence in social media and

common misconceptions about influence and influencers.

(Why) are Online Social Networks so Important?

Lots of Personal Information is Revealed

• An Advertisers Dream!

Online Social Networks and eCommerce

Your Connections Reveal Your Preferences

• “Homophily” – People tend

to interact with others like themselves… “Birds of a

feather flock together.”

• If your friends like hiking, soccer, reading and the

Killers you are more likely

to like hiking, soccer, reading and the Killers.

• In this way – network connections can reveal

your preferences.


• Friends can tell friends how much they like what

they bought or experienced.

• You trust your friends opinions more than a generic

advertisement.

• Example: Facebook’s “Beacon” System

• Trust and privacy must be carefully addressed by

social network based advertising. (Beacon was shut

down and Mark Zuckerberg apologized).

• If we have time (1:27):

http://www.youtube.com/watch?v=1CGF00VIxB8

Peer to Peer Marketing




Predicting Demand (Purchase) from Networks

• Knowing who is connected with whom enables you

to identify the mostly likely adopters of products.

“Network Neighbors Study” (Hill, Provost, Volinsky)*

Marketing to Offline Social Networks

* “Network-based Marketing: Identifying likely adopters via consumer networks. ” Statistical Science 21 (2) 256–276, 2006.

• Product: new communications service

• Long experience with targeted marketing

• Sophisticated segmentation models based on data,

experience, and intuition

– e.g., demographic, geographic, loyalty data

– e.g., experience and intuition regarding the types of

customers known or thought to have affinity for this type

of service

• Added: Whether ‘network neighbors’ (by phone calls)

had already adopted the service.





Prior Adopter

Non Adopter

Network

Neighbor

Adopter

Network

Neighbors

Non-Adopters





1

4.82

2.96

0.4

Non-NN 1-21 NN 1-21 NN 22 NN not

targeted

(0.28%)

(1.35%)

(0.83%)

(0.11%)

Relative Sales Rates for Marketing Segments

Network Effects

Digital Strategy: Session 5

Sources of Positive Feedback

Supply-side economies of scale (Traditional markets)

• More customers � more units produced � lower average

cost per unit

• Marginal cost less than average cost

• Spreading fixed costs across more units

• Manufacturing efficiencies, learning by doing

Demand-side economies of scale (Digital markets)

• More units consumed � higher value per unit

• The value of the good comes from the network of consumers

who use it (at least in part)

• Most commonly caused by network effects (Microsoft,

Playstation, Facebook)

• Positive relationship between popularity and value

Consumer expectations are key!

Virtuous vs. Vicious Cycle

• Expectations matter! Users want to join the

network of winners

• “Rich get richer, poor get poorer”

number of compatible users

value to

user

virtuous

vicious



Markets With Network Effects

• A market exhibits network effects (also known as

“increasing returns to scale” in consumption) when the value to a buyer of an extra unit is higher when more

units are sold, everything else being equal

– A node can reach more nodes in a large network

– Large sales of components of type A induce larger

availability of complementary components B1, ..., Bn,

thereby increasing the value of components of type A

The Model

• Value of a product in a market with

network effects is given by:

Zt is the size of the network at time t,

α represents the value without network effects

γ represents value from network effects.

tZV γα +=

Network Markets: History Matters (I)

� A and B are incompatible but have the same price

� A is available at time 0. B will be available at time t, but

customers do not know its availability until t.

� A and B have intrinsic values of a and b respectively

� Network value is c per user for both products

� Customer arrival rate is 1 per unit time

Network Markets: History Matters (II)

a

b

a+ct

t0

Value

Time

Q: Which product will a new customer at time t adopt? Why?

Network Markets: History Matters (III)

� The superior product, B, is not adopted.

� For network products, both intrinsic performance and installed

base matter.

� A has an inferior performance, but has an installed-base advantage by time t, with total value a+ct>b.

� This is precisely why the inefficient QWERTY keyboard hasn’t

been replaced.

Network Markets: Compatibility Matters

� What happens if B is compatible with A?

a

b

b+ct

t0

Value

Time

Q: What’s the network size of B at time t? Why?

a+ct



Network effects: Tipping

number of users

value to

each user

� More units consumed –> higher value per unit

� Tipping: Success feeds on itself and strong positive feedback can

lead to a “winner-take-all” situation. (eg: Netscape vs. Mosaic, IE vs.

Netscape, Wintel vs. Apple, Nintendo vs. Atari)

� Inferior products that move first may dominate

� Product introduction is difficult, entry strategy is crucial

PRODUCT A

PRODUCT B

Network effects and ‘tippy’ markets

tZV γα +=


PRODUCT A

PRODUCT B

tZV γα +=


PRODUCT A

PRODUCT B


PRODUCT A

PRODUCT B


PRODUCT A

PRODUCT B



PRODUCT A

PRODUCT B


Social Capital in Networks


Value in Structure

• Putting aside personal data, there is valuecreated as a function of network structure.

• Value created for the people or entities in the network:

• How structure generates value for people in the network, (in different network positions)…

“Social Capital”

Social Capital

Social Capital: “[T]he aggregate of the actual or potential resources which are linked to possession of a durable network of … institutionalized relationships of mutual acquaintance or recognition.”

-- Bourdieu (1986: 243)

• Resources?

• Actual or Potential?

• Durable?

• Why include “recognition”?

Why Should We Care?

Individuals in Favorable Network Positions,

those with more “Social Capital”…

• Find better jobs more quickly

• Are more likely to be promoted earlier

• Close deals faster

• Receive higher performance evaluations

• Receive larger bonuses

• Enhance the performance of their teams

• Are more likely to generate innovation (good ideas)

What is a favorable position?

Why is it favorable?

How do benefits accrue to actors in those positions?



What is a Social Network?

� A (Social) Network: A collection of (individuals)

things connected through different types of ties.

(friendship, work relationship, family ties, co-purchase)

� Nodes – People or things that are connected.

� Edges – Links, ties or relationships that connect them.

� Network Structure – The shape of the web of

connections.

Examples of Social Networks

Interdisciplinary collaboration network at the Santa Fe Institute


High-school friendship network


High-school dating network

Notice how the

structure is different for friendship and

dating!

Structure can tell us

a lot about how social groups are

organized, how

information flows, how influence

spreads!


Phone Call Traffic at a Global Media Firm

We Can Map How

Information Flows!


Firm Communication Network

Communication Network:Firm Headquarters

.55 (.15).32 (.20)Researcher

Information

Diversity

Network

Constraint

.55 (.14).30 (.18)Consultant

25.3212.20Clustering

Coefficient

11.02 (32.44)5.41 (19.08)Average

Density

.59 (.12).24 (.14)Partner

.57 (.14).29 (.17)Mean

3473Recruiters

HQFIRM

Firm Communication Network

Communication Network:Firm Headquarters

.55 (.15).32 (.20)Researcher

Information

Diversity

Network

Constraint

.55 (.14).30 (.18)Consultant

25.3212.20Clustering

Coefficient

11.02 (32.44)5.41 (19.08)Average

Density

.59 (.12).24 (.14)Partner

.57 (.14).29 (.17)Mean

3473Recruiters

HQFIRM

Email Network at an Executive Recruiting Firm

Location, Location,

Location – Geography Matters!




Sexual contact network

We Can Model How

Diseases Spread!


Internet Relay Chat (IRC) Channel

Examples of networks

Machine Learning Papers

Networks of Things!

In this case, papers

linked by citations!


The Web, circa 1998

Web Sites Connected

through Hyperlinks


Books linked by co-purchases (we return to this in “The Long Tail”)

Networks Are Dynamic

• http://www.youtube.com/watch?v=8TRzrgMlOKc

• http://www.youtube.com/watch?v=55Q4BwkkRQU



What are “strong ties” “weak ties”?

Strong Ties and Weak Ties

A B C D

“Strong Tie” “Weak Tie”

More Frequent Interaction

Less FrequentInteraction

More EmotionalAffect

Less EmotionalAffect

More Trust Less Trust

More InformationShared

Less InformationShared

Strong Ties and Weak Ties

What is “the strength of weak ties”?

Why are weak ties important?

Building Blocks of Network Structure

A

B

C

“The Forbidden Triad”(Granovetter 1973)

� If A has a strong tie to B, and

� If A has a strong tie to C

=> It is highly likely that B and C

have at least a weak tie.

Why?

“Triadic Closure”

Building Blocks of Network Structure

“Triadic Closure” => “Clustering” in Networks

A

B

C

D

E

F

G

H

I

J

K

l

Clustering

Co-authorship Graph



Clustering

Biotech Collaborations

Clustering

Owen Mundy’s Facebook Graph

Small Worlds Graphs

Watts and Strogatz (1998)

Normal Small Worlds Random

Implications for Degree Distributions

Barabasi and Albert (1999)

Clustering Enables Brokerage

Clustering => Opportunities for “Brokerage”

Bridges are usually Weak Ties

Opportunities for “Brokerage”…

are typically enabled by weak ties

Bridging Weak Ties

span “Structural

Holes”



Who is the Broker?

Burt (2005:14)

S-hole is the mechanism underlying Granovetter’s claim that weak ties are

more useful because they give actors access to nonredundant information

Two Main Benefits to Structural Holes

Information

Control

Early Promotion Profit Margins

Innovation Creativity and Good Ideas Information Advantage

Value of information comes from from its uneven distribution across local network neighborhoods.

Connection to diverse neighborhoods gives access to novel pools of information.

Novel information is valuable due to its local scarcity.

Actors with scarce, novel information can

� broker opportunities, engage in information arbitrage

� use information as a commodity, or

� apply information to problems that are intractable given local information (innovation).



Source of Information Advantages

Constrained Unconstrained

A 40 Year Old Assumption

� Network structure is associated with

productivity and performance.

� Productivity of information workers (Aral et al 2006)

� Productivity of R&D teams (Reagans & Zuckerman 2001)

� Labor Market Outcomes (Montgomery 1991, 1992)

� Wages, Promotion (Burt 1992), Job Placement (Granovetter 1973)

� Innovation (Burt 2004)

� Key theoretical mechanism: access to

information.

DIVERSE

NETWORKS

DIVERSE, NOVEL

INFORMATION

PRODUCTIVITY,

PERFORMANCE,

INNOVATION

The Diversity Bandwidth Tradeoff

Channel BandwidthLowHigh

Network DiversityLow High

A Constrained Network of

Strong, High Bandwidth

Ties

A Diverse Network of

Weak, Low Bandwidth

Ties

Channel BandwidthLowHigh

Network DiversityLow High

A Constrained Network of

Strong, High Bandwidth

Ties

A Diverse Network of

Weak, Low Bandwidth

Ties

Diversity � weak ties => lower bandwidth, frequency, topical dimension, detail, complexity.

The Theory is Problematic because Structural Diversity is likely to be associated with weak

ties. Creating A Tradeoff Between Network Diversity and Channel Bandwidth.

Information Environment Mediates Tradeoff

1. Information Overlap – The degree to which

topics are uniformly or heterogeneously

distributed over nodes.

2. Size of the Topic Space – How many distinct

topics exist in the network.

3. Information Turbulence (the Refresh Rate)

– How often actors’ information is refreshing or

changing per unit time.

Cohesion (The Strength of Strong Ties)

The Value of Cohesive Networks: High trust in a community with

cohesive networks - strong ties fosters mutual assistance obligations

and the social control of deviant behaviors (Coleman 1988)

Ronald Burt: Ego gains numerous competitive advantages and higher

investment returns if ego’s weak, direct-tie relations span structural

holes, thus serving as bridge between alters

Holes create social capital via brokerage opportunities

► Ego actor gains earlier access to flows of valuable information

► Ego fills structural holes by forging new ties linking its unconnected alters,

extract “commission” or “fee” for providing brokerage services

► Low network constraints result in high performance rewards

► Ego maximizes its self-interests by controlling & exploiting information, playing

one actor against another (“tertius gaudens”)

Local vs Global Structural Holes

Reagans and Zuckerman (2001)



Summary

1. Triadic Closure (Forbidden Triad)

2. Clustering and Small Worlds (Watts & Strogatz)

3. Strength of Weak (Bridging) Ties (Granovetter)

4. Structural Holes and Brokerage (Burt)

5. The Diversity Bandwidth Tradeoff (Aral and Van Alstyne)

6. Network Cohesion (Strength of Strong Ties)

(Coleman)

7. Local vs Global Structure (e.g. Reagans and Zuckerman)

Causal Inference, “Influentials” and Network Marketing


@sinanaral @sinanaral

@sinanaral@sinanaral

72 © 2009 Sinan Aral. All rights Reserved.

1. Convince you that :

Success in Social Network Marketing is about

Identifying Causal Peer Influence in Networks.

2. Show you Two Examples of How We Do It

Observational

Experimental

Causality



“Obesity is Contagious”

Christakis & Fowler (2007)



“Di


“We evaluated a densely interconnected social network of 12,067 people assessed repeatedly

from 1971 to 2003 as part of the Framingham Heart Study. The body-mass index was

available for all subjects. We used longitudinal statistical models to examine whether weight

gain in one person was associated with weight gain in his or her friends, siblings, spouse, and

neighbors.”

“[S]tatistical analysis involved the specification of longitudinal logistic-

regression models in which … obesity status at any given time (t+1) was a function of

various attributes, such as the ego’s age, sex, and educational level; the ego’s obesity status

at the previous time point (t); and most pertinent, the alter’s obesity status at times t

and t+1.”


�Obesity Movie



“Di

“CONCLUSION: Network phenomena appear to be relevant to the biologic and behavioral trait of obesity, and obesity appears to spread through social ties.”

“Results: Discernible clusters of obese persons were present in the network at all time points …

A person's chances of becoming obese increased by 57% (95% confidence interval [CI],

6 to 123) if he or she had a friend who became obese in a given interval ... These effects were

not seen among neighbors in the immediate geographic location. Persons of the same sex

had relatively greater influence on each other than those of the opposite sex. The spread of

smoking cessation did not account for the spread of obesity in the network.”



Two Main Arguments Support Claims of Influence

1. Assortative Mixing – Correlation of

Observed Behaviors and Network Structure [e.g. Birke & Belchamber 2009, Christakis & Fowler

2007, 2008, Java et. al. 2006]

2. Temporal Clustering – Friends adoption of

the behavoir is correlated in time [e.g.

Anagnostopoulos et al. 2008, Crandall et. al. 2008,

Christakis & Fowler 2007, 2008]


Logic of the Temporal Clustering Argument…

Anagnostopoulos et. al. [15: 3] argue “if influence does not play a role, even though an agent’s probability of activation could depend on her friends, the timing of such activation should be independent of the timing of other agents.”

Is this convincing?




Research Questions

To what extent do networked social relationships influence

economic decisions (e.g. product adoption), creating

systematic population level patterns in product demand and

other behaviors?

Can we distinguish “influence” in networks from “homophily”

and other confounding factors? If so, what is the relative

importance of each in explaining clustering, and what is the

bias in parameter estimates that do not adequately identify

influence?

@sinanaral

Influence in Social Media Networks

@sinanaral

@sinanaral

@sinanaral

Social Influence:

How the behaviors of one’s peers change

the likelihood that (or extent to which) one

engages in a behavior.

…Behavior change…

Aral, S. (2011) “Identifying Social Influence: A Comment on Opinion Leadership and Social Contagion in New Product Diffusion.” Marketing Science; 30(2): 217-223.

A Stricter Definition of “Influence”@sinanaral

Causal Inference

@sinanaral

The “Reflection Problem”

Human behaviors cluster in

network space and time…

but is this because of

peer influence or

alternate explanations?

@sinanaral

� Homophily (Aral et al. 2009)

� Latent Homophily (Shalizi and Thomas 2011)

� Confounding Factors (Aral and Walker 2011, 2012)

� Simultaneity (Godes and Mayzlin 2004)

� Unobserved Dyanmic Heterogeneity (Van den Bulte and Lilien 2001)

� Truncation (Van den Bulte and Iyengar 2010)

� Other Contextual and Correlated effects (Manski 1993)

Estimation Challenges



@sinanaral

“If you see a crowd of people all put up their umbrellas at the same time, you don’t assume that social influence is responsible.”

– Max Weber

@sinanaral

“Yahoo Study” - Data

� Global IM Network of 27 Million Users from Yahoo! (Daily Traffic)

� Detailed demographics and geographic data.

� Comprehensive, detailed and precise data on online behaviors/activities.

� Day by Day adoption and usage of a mobile service application (Yahoo Go) launched in July 2007. (532,365 Adopters)

Jul 1 Aug 1 Sep 1 Oct 1 Nov 10

2500

5000

7500

10000

12500

Ad

op

ters

pe

r d

ay

Time

Aral, Muchnik & Sundararajan (2009) “Distinguishing Influence Based Contagion from Homophily Driven Diffusion in Dynamic Networks,” PNAS, December.

@sinanaral

0

2

4

6

8

10

12

14

16

20 40 60 80 100 120 140 160

INF

LU

EN

CE

DAYS SINCE LAUNCH

P(Adoption | Friend Adopts)

Controlling for correlated preferencesand confounding effects

The “iPad Effect”

@sinanaral

Homophily Exaggerated Among Early Adopters

Jul Aug Sep Oct

0.7

0.8

0.9

Adopter friends

Non-adopter friends

Random user

Co

sin

e D

ista

nc

e

Time

Sim

ilari

ty

Time

@sinanaral@sinanaral @sinanaral

Constructed Observational Evaluation(Eckles 2012; Bakshy et al 2011)

1. HDPSM can achieve 80% Error Reduction.

2. Context Relevant Variables are Key.



“VIRAL DESIGN” Can we engineer products so

they are more likely to be virally shared?

@sinanaral

Aral & Walker (2011) “Creating Social Contagion through Viral Product Design: A Randomized Trial of Influence in Networks,” Management Science, September.

@sinanaral

The Setup

App

� Randomly Enabled Viral Messaging.� Observed the Adoption and Use of the App by Friends of

Control and Experimental Group Users.

@sinanaral

Data

~ 10K Experimental Users

~ 1.4M Friends of Experimental Users

We Observe Application Diffusion Over this Network

1. Facebook Profiles

2. Adoption

3. Use

Flixster - An Example Facebook Application

Flixster - An Example Facebook Application

Users can invite their friends

to adopt the application and join their social network on

the application itself.









Users can invite friends to install this application and

include a personal message







Invites are a form of Viral Messaging

Another form of viral messaging is

notifications

Notifications are generated

automatically when a user takes an

action within an application. They

are delivered to a user’s Facebook

friends like this

In addition to viral messaging, Facebook applications also make use of traditional

online advertisements by placing ads directly inside the application region.

There is a market for Within-Application

advertising.

@sinanaral

� Leakage and contamination could occur if peers are

a) connected through indirect pathways,

b) connected to multiple treated peers in different treatment groups or

c) connected to multiple treated peers in the same treatment group.

� These scenarios are rare in our data

� We control for leakage and peers of multiple treated users by only evaluating recruited users and right censoring contaminated peers.

� Contaminated: Any peer with multiple treated peers after time t at

which they have multiple treated peers.

� Sensitivity analysis shows censoring has little effect. This may

make our results more conservative but minimizes contamination.

Contamination, Leakage & Interference



@sinanaral

“Inside-Out” Estimation

βλλ kiX

kkik etXt )(),( 0=

Variance Corrected Stratified Duration Models

),()0|1( 1 jtj ijititit ywxFyyP ∑=== − βγ

Conventional Approach in Observational Data

@sinanaral

Personal

Invitations

Passive

Awareness

Influence Per

Message

Global Diffusion

Stickiness

↑6% ↑2%

↑98% ↑246%

↑17% 0%

Which Features Spread Contagion Best?

Viral

Invitation

Viral

Notification

Email

Campaign

FB

Banner

Web

Banner

CTR

Relative Marketing Effectiveness

6% 2%2%-5.9%

.10%-.20%

.07%

@sinanaral

@sinanaral

@thesocialcure



@sinanaral

• Research in 1950’s emphasized importance of personal influence

– Trusted ties more important than media

influence in determining individual opinions

• Also found that not all people are equally influential

– A minority of “opinion leaders” or “influencers” are responsible for influencing

everyone else

• Called this “the two-step flow” of information

– “One in ten Americans tells the other nine how to vote, where to eat, and what to buy.” (Keller and Berry, 2003)

The Two-Step Flow, Opinion Leaders, & Influencers

@sinanaral

The “Law of the Few”

• Since 1950’s idea that minority of special people has a vastly disproportionate impact on social change has caught on

– Gladwell (2000) called this “the law of the few”

• Not really what K&L were saying

– Two-step flow says on that influencers decide which

information to pass on

– Law of the few says that they trigger social epidemics

• Nevertheless, this idea has become extremely popular

@sinanaral

=

“Social epidemics ... are also driven by the efforts of a handful of exceptional people” Gladwell (2000)

IT’S SUCH A GOOD STORY…

@sinanaral

Individuals work like this…Influentials Hypothesis assumes

society works the same way…

IT MAKES SENSE…

@sinanaral

Marketers Love Influencers

“Influencers have become the ‘holy

grail’ for today’s marketers.”

—Rand (2004)

@sinanaral

BUT GRAILS ARE HARD TO FIND…



@sinanaral

Experiment 2

We also Randomized Receipt of Notifications

Only a Randomly Selected Subset of Neighbors

Receive Passive Viral Messages

Identifying “Influentials” and “Susceptibles”

Allows us to test:

Randomized Trails of Influence and Susceptibility to Influence

Aral, S. & Walker, D. (2012) “Identifying Influential and Susceptible Members of Social Networks.” Science; 337 (6092): 337-341.

@sinanaral

Results

� Influence increases with Age

� Susceptibility decreases with Age

� Women are less susceptible to influence than men

@sinanaral

Results

� Influence transmits over relationship pairs of the same age.

� Suggestive evidence that older people influence younger people more than younger people influencing older people.

� Women are less susceptible to influence than men

� Influence transmits more relationship pairs where the sender is of the same or greater level of relationship commitment as the recipient.

Susceptibility to Influence

1

1.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

2

Single Relationship Engaged Married Its Complicated


Results




@sinanaral

Learning Objectives: Network Analytics

1. Understand why (Social) Networks are so critical to demand prediction and marketing.

2. Understand Economic Network Effects and why they are

so essential to Digital Strategy.

3. Understand the importance of Causal Statistical

Estimation in effective Social Network Marketing efforts.

4. Understand Viral Product Design and its implications for

a) Social Contagion in Product Adoption, b) Sustained

Product Use and c) the relationship between the two.

5. Consider how to identify influence in social media and

common misconceptions about influence and influencers.

nyu stern spring 2013 - mitweb.mit.edu/sinana/www/digital analytics & strategy 2013 -...

Documents