| Notes |
Social Physics and
the Data Driven Society
Prof. Alex Pentland, MIT
Connection Science and Engineering
WEF Big Data, Hyperconnected World
Outline
1. Big Data and computational
social science
2. Distributed Intelligence
3. Network Intelligence
4. Big Data breaks science
Copyright alex pentland 2011
1. Understanding Ourselves:
The Big Data Revolution
Human Dynamics Observatories:
(1) MIT Reality Mining Study
(2) MIT Social Evolution,
(3) MIT Friends and Family (Current),
(4) MIT lifelog pioneers; MyLifeBits,
(5) Sociometric Badge studies,
(6) Midwest Field Station ,
(7) Framingham Heart Study,
(8) Large Call Record Datasets ,
(9) “Omniscient”/All-Seeing View
Ahrony, Pentland
Background:
Humans Have Two Types of Thought
Nobel Prize winner Kahneman,
father of behavioral economics
Fast
Parallel
Automatic
Associative
Slow
Serial
Controlled
Rule-based
Habitual
(System 1)
Attentive
(system 2)
Conceptual Representations
Past, Present, Future
Can be evoked by language
Content Process
Social
Physics
Copyright alex pentland 2012, all rights reserved
People Mostly Learn by Examples,
not Arguments or Reasoning
90% - 10% balance
Rendell et al, Social Learning, Science 4/10
Copyright alex pentland 2012 all rights reserved
6
Influence Model & Idea Flow
s1
s2
s1 0.8 0.2
s2 0.2 0.8 Internal State s1
Internal State s2
Internal State s2
Private
Observation
Copyright alex pentland 2012, all rights reserved
7
Using sensors in
smart phones to
obtain different type
of social networks.
Call Network
Proximity Network
Friendship Network
Co-location Network
Copyright alex pentland 2012, all rights reserved
45% accuracy predicting
app downloads. Gompertz
function describes influence
Social Exposure Predicts Behavior
Pan, Aharony, Pentland
65 young families, 12 months data
8
process id process id
(c)
1 2 3 4 5 6
1
2
3
4
5
6
Inverse Problem:
Discovery of Influence, Node State
Raw Observations From Nodes
clean
state
estimate
for nodes
influence
structure
connecting
nodes
Copyright alex pentland 2012 all rights reserved
Understanding Ourselves:
Behavioral Demographics
Copyright alex pentland 2012
90 million people continuously
Accuracy 4 times
normal demographics
Copyright 2012 Alex Pentland
Patterns of Health
With MGH: Phenotypic + Genetic Characterization
Alcohol
diabetes
Copyright 2012 Alex Pentland
Patterns of Finance
Success Scoring of Unbanked
Unlikely to
succeed
Likely to
succeed
Including life coaching
2. Distributed Intelligence
Shaping By Social Incentives
40th Anniversary of the Internet Grand Challenge
Copyright alex pentland 2012 Pickard, Pan, Rahwan,Cebrian,Madan,Crane,Pentland
Shaping by Social Incentives
incentives that leverage social influence
Global externality:
tragedy of the commons
Localized externality:
The peers of individuals A and B
receive rewards for behavior of
A, B.
Mani, Rahwan, Pentland Copyright alex pentland 2012 all rights reserved
Behavior Shaping By
Social Influence
• Reward individuals for their peers' behavior
• The total reward distributed to the peers of actor j is
less than the Pigouvian subsidies to j if
Mani, Rahwan, Pentland
Incentive
personal
utility
externality
cost
peer
pressure
peer
cost
Copyright alex pentland 2012 all rights reserved
Social Influence incentive mechanism is 3.5 times
as efficient as standard incentive mechanism
65 young families, 3 months data
Aharony, Pan, Pentland
Standard
Incentive
Social media
Incentive
Peer reward
incentive
Copyright alex pentland 2012 all rights reserved
3. Network Intelligence
•Learn the underlying hidden influence network from historical data
•Use edge weights (of network) to derive adoption potential
Influence Model
Influence Model
: edge weight between i and j in the diffusion network
= 1 if j has adopted strategy a, =0 if not.
Copyright alex pentland 2012 all rights reserved Altshuler, Pentland
•Learn the underlying hidden influence network from historical data
•Use edge weights (of network) to derive adoption potential
•Calculate behavior predictions
Behavior Change Model
Behavior Propagation
Diffusion model:
is the individual susceptibility factor of user u
a s p u
u
u
Local
u a
P E Prob x 1| Ν u 1 e
Copyright alex pentland 2012 all rights reserved Altshuler, Pentland
•Learn the underlying hidden influence network from historical data
•Use edge weights (of network) to derive adoption potential
•Calculate behavior predictions
•Predict cascade frequency and size, from local influence forces
Model
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Trend Penetration Probability
Probability of Idea-Behavior Flow, Φ(C)
Copyright alex pentland 2012 all rights reserved
Social trading: users
can see and copy
trades of another user
Φ(C) and the
Wisdom of the Crowd
Copyright 2012 Alex Pentland All Rights Reserved Altshuler, Pentland
Isolation and Herding
User IDs
User IDs
•2.7 Million users
•“Twitter-like” social based financial trading
•Trading as collaborative problem solving
eToro – Social Trading Network
Copyright alex pentland 2012 all rights reserved
Altshuler,Pentland
Social Trading (Annual ROI)
Social Trading
Non-social trading
0 50 100 150 200 250 300 350 400
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
Days
Cummulative daily ROI
Combined, Max drop: -7.81% , ROI: 5.58% , ROI\ STD: 22.25 , Days win ratio: 2.34
Social Trading
ROI : 5.58%
Max drawdown : -7.81%
Sharpe (yearly) : 1.03
Days win ratio : 0.7
Social Intelligence
confidential Athena Wisdom 2012` Altshuler,Pentland
Stupidity of the Crowd
Copyright 2012 Alex Pentland all rights reserved Pan, Alshuler, Pentland
“Guru trading”
Social Trading
Non-social trading
0 50 100 150 200 250 300 350 400
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
Days
Cummulative daily ROI
Combined, Max drop: -14.49% , ROI: 18.22% , ROI\ STD: 32.56 , Days win ratio: 2.35
“Guru Trading”
ROI : 18.22%
Max drawdown : -14.49%
Sharpe (yearly) : 1.56
Days win ratio : 0.7
Guru Trading (Annual ROI)
Experts
confidential Athena Wisdom 2012 Altshuler,Pentland
Copyright alex pentland 2012 Return on Investment Social vs non
-social investors
ROI for 2.7 Million Investors, 1 Year
Isolated
Echo
chamber
Idea->Behavior Flow, Φ(C)
Decision Accuracy Depends on
Diversity of Information Sources
Altshuler,Pentland
Selecting Oracles
Confidentiall Athena Wisdom 2012
0 5 0 100 150 200 250 300 350 400 450 500
-50%
0
50%
100%
150%
200%
250%
Trading days ROI
ROI: 209.4%
Sharpe: 6.1
Drawdown: - 4.9%
Daily win ratio: 85.3%
Altshuler,Pentland
Insuring Diversity of Information
User IDs
User IDs
•2.7 Million users
•“Twitter-like” social based financial trading
•Trading as collaborative problem solving
eToro – Social Trading Network
Copyright alex pentland 2012 all rights reserved Altshuler,Pentland
Tune Network to Optimize Φ(C)
Day of the experiment
ROI in % (Social
– Non-Social)
Currency Trading ROI
Φ(C)
Copyright alex pentland 2012 all rights reserved Altshuler,Pentland
Pattern of Social Ties and Φ(C)
Engagement:
Density of sharing
of information
within group
Exploration:
Harvesting New Ideas
outside of group;
`fat tails’
Copyright alex pentland 2012 all rights reserved
Copyright 2012 Alex Pentland All Rights Reserved
Exploration and Engagement:
a study of white collar workers
Engagement in face-to-face accounts for
30% of between-group variation in productivity
Exploration in face-to-face accounts for
10% of between-group variation in productivity
Wu,Waber,Aral,Brinjolfsson,Pentland
Best Research Paper, ICIS 2008
Copyright alex pentland 2012 all rights reserved
BAC Call Center
Productivity Study
Average Call Handle Time
Engagement:f Face-to-Face Network
Productivity correlated
with group engagement
Phase 1
Phase 2
Optimize Idea Flow Φ(C)
Energy
Engagement
Average Call Handling Time
Copyright alex pentland 2012 all rights reserved Olguin, Waber, Kim, Pentland
Changing coffee break
structure produced:
30% increase engagement
20% decrease stress
$15M / year savings
Copyright alex pentland 2012 all rights reserved
Copyright 2012 Alex Pentland All Rights Reserved Olguin, Waber, Kim, Gloor, Pentland
Φ(C) Measures Are Typically 40%
of Performance
Harvard Business Review:
Breakthrough Idea of the Year
Cities and Φ(C)
Lieben-Nowell; Krings et al;
Engagement:
Density of sharing
of information
within group
Exploration:
Harvesting New Ideas
outside of group
Copyright alex pentland 2012 all rights reserved
Φ(G) Φ(G)
Φ(C)
4. Big Data breaks science
Science as practiced assumes
strong theoretical understanding
Big Data is good for interpolation
but not for extrapolation
Big Data governance requires
thousands of social science
experiments
Trento, Italy
Copyright alex pentland 2012, all rights reserved
Data from private companies and Provential Authority
Trentino Open Living Data Project (TOLD)
Application scenarios:
Mobility:
• Online efficient private traffic
• Public transportation on the fly route
balancing
Safety:
• Detection and support in dangerous
situations (e.g. fires, avalanches, etc.)
Health:
• Recognition and prediction of
epidemic spread
Urban & Local business planning:
• Understand economically depressed
areas
• Help companies to plan investment
• A joint project between
Copyright alex pentland 2012, all rights reserved
Data from individuals
Mobile Territorial Lab • Understand the needs and the
behaviour of users.
• Provide individuals mobile phone
equipped with a sensing middleware to
collect the data generated to be
analyzed (starting community: young
families with newborns)
• Short term outcomes:
1. Developing and testing a new model
of DATA OWNERSHIP
2. Understanding the dynamics of
people’s needs
1. Understanding people’s interactions
in the generated social networks
A joint project between:
Copyright alex pentland 2012, all rights reserved
Copyright alex pentland 2012, all rights reserved
Forbes, 8/10,
Mining Human Behavior
pentland@mit.edu
http://media.mit.edu/~pentland |