online actions with offline impacttimalthoff.com/docs/althoff-2017-social-influence... ·...
TRANSCRIPT
Online Actions with Offline Impact: !How Online Social Networks Influence Online and Offline User Behavior
Tim Althoff, Pranav Jindal, Jure Leskovec @timalthoff
Alice Bob
Tim Althoff, Stanford University
Alice Bob
Tim Althoff, Stanford University
What happens to Alice & Bob?
Alice Bob
Tim Althoff, Stanford University
friend
Previous Research: Online Behaviors
3
Online Behavior
Tim Althoff, Stanford University
§ Homophily – user similarity § Triadic closure – common friends § Community structure § Network growth § …
Want to understand online à offline!
4
This talk
§ Homophily – user similarity § Triadic closure – common friends § Community structure § Network growth § …
Online Behavior
Offline Behavior
Tim Althoff, Stanford University
?
§ Physical activity § Diet § Smoking § Political mobilization § …
What causes change in behavior?
5 Tim Althoff, Stanford University
Alice
’s Ac
tivity
Time
What causes change in behavior?
5 Tim Althoff, Stanford University
Alice
’s Ac
tivity
Time
What causes change in behavior?
5 Tim Althoff, Stanford University
Alice
’s Ac
tivity
Time
Why !increase?
What causes change in behavior?
5
Social Influence
Tim Althoff, Stanford University
Alice
’s Ac
tivity
Time
Why !increase?
What causes change in behavior?
5
Social Influence
Intrinsic Motivation
Tim Althoff, Stanford University
Alice
’s Ac
tivity
Time
Why !increase?
Main Research Question § Do online social networks influence
offline (real-world) behavior? § Challenge: How to disentangle social
influence from intrinsic motivation? § Never observe alternate universe where
the friendship did not happen
6 Tim Althoff, Stanford University
Our Contributions 1. New method for causal identification of
social influence § A natural experiment exploiting delay of
friendship acceptance
2. Social networking leads to a 5 months !long increase of up to ~350 daily steps
3. Can predict which users will be most
influenced by new social network connections with 79% accuracy
7 Tim Althoff, Stanford University
Activity Tracking Dataset § Azumio’s Argus activity
tracking app (iOS/Android)
§ 6 million users § 211k users join social network § Enables interaction & notifications
§ 791 million actions § online (posts) § offline actions (physical activity; !
accelerometer-based) § 5 year timespan
8 Tim Althoff, Stanford University
Disentangling Motivation & Influence
9 Tim Althoff, Stanford University
Disentangling Motivation & Influence
9 Tim Althoff, Stanford University
Disentangling Motivation & Influence
9
Direct accept Delayed accept
Tim Althoff, Stanford University
Disentangling Motivation & Influence
9
Direct accept Delayed accept
Tim Althoff, Stanford University
Disentangling Motivation & Influence
9
Direct accept Delayed accept
Difference isolates social influence
influence
motivation motivation
influence
Tim Althoff, Stanford University
A natural experiment for Alice? § Is assignment to “direct accept” (<1 day)
and “delayed accept” (>7 days) !as if random?
à Groups would be indistinguishable! or balanced
§ Balancing criterion: Covariates within standardized mean difference (SMD) of 0.25 SD [Stuart, 2010]
10 Tim Althoff, Stanford University
Testing Alternative Hypotheses Maybe direct accepter’s are… § …more active?
§ Avg. daily steps 7 days before § 0.07 SD < 0.25 SD à No
§ …closer friends? § # common friends at request § 0.08 SD < 0.25 SD à No
§ …newer to social network? § # days since joining § 0.179 SD < 0.25 SD à No
+19 further hypotheses in paper; all balanced
11 Tim Althoff, Stanford University
Testing Alternative Hypotheses Maybe direct accepter’s are… § …more active?
§ Avg. daily steps 7 days before § 0.07 SD < 0.25 SD à No
§ …closer friends? § # common friends at request § 0.08 SD < 0.25 SD à No
§ …newer to social network? § # days since joining § 0.179 SD < 0.25 SD à No
+19 further hypotheses in paper; all balanced
11
“True” natural experiment!
Tim Althoff, Stanford University
Experimental Results
Delayed acceptedrequest
Acceptedrequest
0
50
100
150
200
250
300
350
400St
eps
diffe
renc
e
} }}M
M + I
I
12 Tim Althoff, Stanford University
45%
55%
Experimental Results
Delayed acceptedrequest
Acceptedrequest
0
50
100
150
200
250
300
350
400St
eps
diffe
renc
e
} }}M
M + I
I
12 Tim Althoff, Stanford University
45%
55%
Experimental Results
Delayed acceptedrequest
Acceptedrequest
0
50
100
150
200
250
300
350
400St
eps
diffe
renc
e
} }}M
M + I
I
12 Tim Althoff, Stanford University
45%
55%
Causal effect of increasing !
real-world activity !due to online social network
by ~350 steps/day !
Effect of Social Network Over Time
4 3 2 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 207ime in weeks
5600
5800
6000
6200
6400
6600
Average daily steps 7reatment
Control
!
13 Tim Althoff, Stanford University
§ Longitudinal study with Difference-in-Differences design and matched control group
Effect of Social Network Over Time
4 3 2 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 207ime in weeks
5600
5800
6000
6200
6400
6600
Average daily steps 7reatment
Control
§ Online social networks significantly increase offline physical activity for ~5 months
13 Tim Althoff, Stanford University
§ Longitudinal study with Difference-in-Differences design and matched control group
Heterogeneous Effects § Robust: Several additional experiments all
reveal similar effects § Joining network also affects online behavior
§ App usage (increased 30%) § One-year retention (increased 17%)
14 Tim Althoff, Stanford University
Heterogeneous Effects § Robust: Several additional experiments all
reveal similar effects § Joining network also affects online behavior
§ App usage (increased 30%) § One-year retention (increased 17%)
§ Repeatability: 1st friend > 2nd friend § Tie Strength: friend > follower § Initiator: sender > receiver § Demographics: > (52% more!)
14 Tim Althoff, Stanford University
Heterogeneous Effects § Robust: Several additional experiments all
reveal similar effects § Joining network also affects online behavior
§ App usage (increased 30%) § One-year retention (increased 17%)
§ Repeatability: 1st friend > 2nd friend § Tie Strength: friend > follower § Initiator: sender > receiver § Demographics: > (52% more!)
14 Tim Althoff, Stanford University
Can use these insights to predict behavior change!
Predicting Behavior Change § Which individuals will be most influenced
by new social network connections? § Binary prediction task: Will you increase
your activity over next 7 days? (55.4% do) § Short-term behavior change
§ Dataset: 432k new friendships § Gradient Boosted Tree models
15 Tim Althoff, Stanford University
Predicting Behavior Change § Which individuals will be most influenced
by new social network connections? § Binary prediction task: Will you increase
your activity over next 7 days? (55.4% do) § Short-term behavior change
§ Dataset: 432k new friendships § Gradient Boosted Tree models
15 Tim Althoff, Stanford University
Can predict positive behavior change with 79% ROC AUC
Summary & Conclusions Findings § Social networks influence behavior – online & offline § Can predict which users will be most influenced by
new social network connections
Implications § How to design & maintain social networks as
effective support communities that lead to healthier lifestyles? § Group composition § Friendship recommendation algorithms § Enabling supportive and encouraging interactions
16 Tim Althoff, Stanford University
Acknowledgements § We thank Azumio for donating data from their Argus
app for independent research.
@timalthoff [email protected] www.timalthoff.com
Ask me anything!
Tim Althoff, Stanford University 17
Travel Award