From Conversations to Conversions: Identifying Peer Influence and Social Contagion in Networks The Rise of Social Commerce October 6, 2010
Sinan Aral NYU, MIT & Social Amp sinan@stern.nyu.edu
Agenda 1. Convince you that : Success in Social Commerce is about Identifying Causal Peer Influence in Networks.
2. Show you Two Examples of How We Do It One Observational One Experimental
The Importance of Social Networks
Interest in Social Networks has exploded...
Many organizations are attempting to collect and analyze large scale social network data…
Call networks
Affiliation networks
Email networks
Friendship networks
Terrorist networks
Organizational networks
New unprecedented data allow us to tackle previously intractable research questions (Lazer et. al. 2009)
A Program of Research To mine massive social network data to better understand:
How behavioral contagions spread in human social networks How peers influence each other to buy products, be more productive, vote for a different political candidate... Â Product Adoption & Demand
 Disease Prevention (Condom Use)
 Productivity of Information Workers
 Market Trends (Buzz)
 Viral Peer-to-Peer Marketing
 Economic Development
Causal Statistical Estimation
The “Reflection Problem”
Manski (1993)
Identifying Causal Peer Effects in Networks is Notoriously Difficult
Now lots of empirical evidence that human behaviors tend to cluster in network space and time,… but is this because of peer influence or alternate explanations?
“Obesity is Contagious”
Christakis & Fowler (2007)
But Homophily can also explain such evidence A long line of theory indicates: • “Birds of a feather, flock together” – attributed to Robert Burton (1577-1640) Recently: Lazersfeld and Merton (1954) and Blau (1977) among others.
• (People) love those who are like themselves -- Aristotle, Rhetoric and Nichomachean Ethics • Similarity begets friendship -- Plato, Phaedrus • Hanging out with a bad crowd will get you into trouble -- Sinan’s Mom
Peer Influence is Critical to Social Commerce
We commonly assume: 1. *Social* Commerce is More Engaging than Commerce (Consumers shop for and are persuaded by their friends) 2. Observing *Social* Commerce Provides More Intelligence (Data on Social Shopping enables Targeting and Referral Strategies that provide demonstrable Sales Lift)
These assumptions necessitate causal evidence of peer influence in social commerce outcomes.
Consider the Implications of Homophily vs Influence for Marketing Strategy…
Target “Influentials” vs. Segment the Market based on consumer profiles.
Dynamic Matched Sample Estimation Aral, Muchnik and Sundararajan (2009)
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 for 5 months. 12500
532,365 Total Adopters
Adopters per day
10000 7500 5000 2500 0 Jul 1
Aug 1
Sep 1
Oct 1
Nov 1
Observe Clustering in Longitudinal Data
Distinguishing Influence from Homophily “Influence” Estimates Comparing Adoption in Treated and Untreated Cases Under Randomized Matching Over Time (Methods used by those who take AM as evidence of influence)
“Influence” Estimates Comparing Adoption in Treated and Untreated Cases In Our Dynamic Matched Sampling Framework Over Time
Much of the estimated influence is really observable homophily.
Viral Product Design Can firms explicitly engineer products so they are more likely to be shared amongst peers? Viral Product Characteristics
Viral Product Features
Content likely to inspire viral sharing
Modalities of use
Usefulness, topicality, prominence, positive valence and unexpectedness (Berger and Milkman 2009, Stephen and Berger 2009, Berger and Heath 2005, Phelps et al 2004, Heath, Bell and Sternberg 2001).
Invites Notifications Hypertext Embedding (No literature)
Targeted Messaging Active-Personalized Messaging is thought to be more persuasive and thus more effective.
People tend to activate their strong ties. Strong ties exhibit greater homophily. Greater pressure for conformity. Deeper knowledge of one another’s preferences. We tend to trust information from close trusted sources. Reciprocity. Personalized messages have been shown to be more persuasive and thus more effective.
But, these messaging channels require more effort and time which may curtail their use. Passive-Broadcast Messaging may reach more people but may be less persuasive. Which is more effective is an empirical question.
Viral Feature Space PERSONALIZATION INCREASING BROADCAST
ACTIVE
PERSONALIZED Peer Network
High Effort Minimal Effort
PASSIVE
No Effort
ACTIVITY INCREASING
Population
Specific Individuals Personalized Referrals
Generalized Hypertext Embedding
Personalized Hypertext Embedding
Automated Broadcast Notifications
Greater marginal peer influence per message
Collaborative Bookmarking
Automated Targeted Notifications
More messages generated
Randomized Experiments on Facebook Facebook is an ideal experimental environment in which to study peer influence: Out in the ‘real’ online world, not in a laboratory. Hundreds of millions of individuals interacting: 500M+ users (35M+ login daily), tens of billions of relationships.
Detailed digital records of user’s online representation and interactions Demographics, preferences, views and interests, social behavior (group participation, communications, status updates), relationships, product preferences, product adoption outcomes
Experimental control is possible through the use of multiple Experimental Facebook Applications developed in concert
The Setup App
Randomly Enabled Viral Messaging. Observed the Adoption and Use of the App by Friends of Control and Experimental Group Users.
Data
Profile and Network Data on ~10 Million Users Profiles are Dynamic – we observe changes weekly
~ 10K Experimental Users ~ 1.4M Friends of Experimental Users We Observe Application Diffusion Over this Network Adoption Use
Experiment: Randomized Viral Messaging We designed multiple experimental versions of a Facebook Application that has been widely advertized and adopted. As users adopted the application…
Experiment: Randomized Viral Messaging We designed multiple experimental versions of a Facebook Application that has been widely advertized and adopted. As users adopted the application…
… we randomly assigned them to Control and Experimental groups.
Experiment: Randomized Viral Messaging We designed multiple experimental versions of a Facebook Application that has been widely advertized and adopted.
‌ we randomly assigned them to Control and Experimental groups.
Experiment: Randomized Viral Messaging We designed multiple experimental versions of a Facebook Application that has been widely advertized and adopted.
‌ and collected data on their personal attributes and preferences from their Facebook profiles‌
Experiment: Randomized Viral Messaging We designed multiple experimental versions of a Facebook Application that has been widely advertized and adopted.
‌ as well as data on their social networks and the personal attributes and preferences of their network neighbors.
Experiment: Randomized Viral Messaging Experimental Group
Control Group
Experimental Group
Viral Messaging Enabled Experimental Users Send Active and Passive Viral Messages to their Neighbors
Control Group Viral Messaging Disabled
Experiment: Randomized Viral Messaging Experimental Group
Control Group
Experimental Group
We also Randomized Receipt of Passive Viral Messages Only a Randomly Selected Subset of Neighbors Receive Passive Viral Messages
Control Group Viral Messaging Disabled
Experiment: Randomized Viral Messaging
We then compare:
Control Group
Experimental Group
9click throughs 9adoption 9usage data of neighbors of: 9Experimental Group 9Control Group
Experiment: Randomized Viral Messaging
We then compare:
Control Group
Experimental Group
9click throughs 9adoption 9usage data of neighbors of: 9Experimental Group 9Control Group
Allows us to test: 1. Average Treatment Effect of Viral Messaging Capabilities on Peer Adoption and Network Propagation 2.Randomized Trails of Susceptibility to Peer Influence via Viral Messaging
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 In addition to viral messaging, Facebook applications also make use of traditionalviral messaging is notifications online advertisements by placing ads directly inside the application region. There is a marketNotifications for Within-Application are advertising. generated automatically when a user takes an action within an application. They are delivered to a user’s Facebook friends like this
Conventional Approach in Observational Data
P ( yit = 1 | yit −1 = 0) = F ( xit γ , β ∑ j wij y jt ) “Inside-Out” Estimation
A Variance Corrected Stratified Proportional Hazards Model
λk (t , X ki ) = λ0 k (t )e X
ki β
Adoption Diffusion Differs Significantly
Passive
Active
Marginal Influence Per Message
2%
6%
Global Adoption Diffusion
246%
98%
Use and Engagement
Network Externalities: 33% Increase in Use Per Friend Referred But: Only Active Recruitment Creates Sustained Engagement
Susceptibility to Influence
Thank You!
From Conversations to Conversions: Identifying Peer Influence and Social Contagion in Networks Sinan Aral NYU Stern School of Business sinan@stern.nyu.edu