ARO MURI - the School of Engineering and Applied Science

Report
W911NF-12-1-0509
ARO MURI: Evolution of Cultural
Norms and Dynamics of
Socio-Political Change
Ali Jadbabaie
Alfred Fitler Moore Professor of Network Science
University of Pennsylvania
The Team
Ali Jadbabaie (PI) Michael Kearns Daron Acemoglu Asu Ozdaglar Munzer Dahleh Fotini Christia
University of Pennsylvania
Matt Jackson Jure Leskovec
Stanford University
Massachusetts Institute of Technology
Jeff Shamma
Georgia Institute of
Technology
Jon Kleinberg
Larry Blume
Cornell University
Motivation and Overview
Goal: create a research program that leads to understanding of social norms,
political change, cultural dynamics, societal stability with a multidisciplinary lens
that leverages network science, systems theory, dynamics, Economics, Political
Economy, Computer Science
• Many of the central questions involve interactions among individuals and groups
with different identities
– Study of collective phenomena and collective decision making in networked setting
with domain specific knowledge
– Need more quantitative approaches, beyond descriptive analysis
– Need theory, principled modeling, data analysis, lab experiments, and field
surveys
– Need to educate a new breed of computational social scientists and
engineers
Why us?
Meme tracker
Synergies between thrusts
Thrust 3:Micro foundations of
social/political interaction
Strategic and behavioral models
(diffusion, contagion, cascade)
Social and political change
Thrust 1:
Evolutions of Social
norms/Social learning/trust
Conceptualizing/
Modeling
Social& Political
Change
Thrust 4:foundations of
Control of sociopolitical
change/contagion
Thrust 2
Political change
How does it all come together?
Network
Science
Jackson, Blume,
Acemoglu
Jackson, Jadbabaie
Blume,Jackson
Economics/
Political
Economy
Systems
Theory
Computer
Science
Experiments/
Field studies
Acemoglu, Christia,
Leskovec, Jadbabaie,
Dahleh
Leskovec, Kearns,
Ozdaglar
Kleinberg, Ozdaglar,
Acemoglu, Leskovec,
Dahleh,Jadbabie,
Kearns,Blume
Shamma, Dahleh
Ozdaglar, Leskovec,
Kleinberg, Jadbabaie,
Shamma
Jadbabaie
Collective behavior,
social aggregation,
dynamics of cascades
Acemoglu
Ozdaglar
Dynamics of
Game Theory,
sociopolitical
Networks,
change, learning Cascades
Jackson
Control, Decision
making, Global
networked games
Social and
Econometrics of
economic networks,social networks,
evolution of social Emergence of trust
norms
• First principles
• Rigorous math
• Algorithms
• Proofs
Modeling
Blume
Data
Analysis
Kleinberg
Christia
Field studies, large Networks, games
Behavioral
Experiments, randomized surveys algorithms,
from conflict zones Modeling cascades
contagion
Dahleh
Theory
Kearns
Leskovec
Shamma
Social networks
data and
experiments
Lab
Experiments
Learning in games,
robustness,
evolutionary dynamics
Real-World
Surveys
• Analysis of social
• Economics
• Stylized, Controlled • Extremely challenging!
•Political Science network data
• Clean, real-world • Randomized, large scale
•Empirical data • Statistics/Econom
data
studies
• How to deal with etrics
“no physics”
Thrust 1: Social Norms/Social
Learning/emergence of trust
• Task S1: Evolution of social norms
– How do social norms such as cooperation and reciprocal altruism emerge and
evolve, How does law abiding behavior emerge, and how are norms related to
laws? When can imposing laws on behavior induce good behavior?
• Key Questions answered via game theoretic model (Acemoglu &
Jackson 2013 ,2014). Key insight History serves as guide:
prominent ``good’’ behavior can induce good behavior, but Good
behavior erodes over time
This year S1,S3: Blume challenges the reciprocal altruism of Axelrod
on emergence of cooperation using a stochastic game. Key
insights:
– Sharing can be costly
– Benefits to cooperation can not be always reciprocated when
individuals/societies are resource-poor. Failure of sharing can be
“contagious”
Thrust 1: Social Learning, Coordination,
Beauty contest games, and Conventions
Tasks S1,S2: Information aggregation, social learning, and
coordination on conventions
Jadbabaie (2013,2014), Jackson 2014: How do individuals in a society
collectively aggregate dispersed information? How does the role of Network
structure and quality of private information affect aggregation, learning and
diffusion? Answer: Interplay of centrality of informed individuals and
discriminative power of their private observations
• Supporting Theory and empirical evidence in Jackson’s talk
on Thrust 3, and Jadbabaie’s talk on Thrust 1
• Jadbabaie (2014): How does preference for convention
affect strategic coordination? Will consensus in actions result in
consensus in estimates?
Jackson(2014): Bayesian learning with limited memory
• Over-interpretation of ambiguous signals causes
polarization.
• Ongoing work (Jadbabaie): create non-Bayesian
updates that are scalable by limiting “recall”
Shamma 2013
Thrust 2: Political Change, development, societal stability
Columbia, Egypt, Afghanistan, Yemen, Sierra leone
• Tasks P1,P2:This year: Power of state/Provision of goods
• Network approach to modeling and estimating the direct and spillover effects of
state capacity on public goods and prosperity using data across Colombian
municipalities.
• Question: what is the effect of “infrastructural” power of the state (e.g.,
presence of the state and its employees) on economic development?
– In many instances, what matters is the local state.
• Presence of the local state, may create spillovers on neighboring areas.
Lawlessness in a municipality will spillover into neighboring areas; efficient
provision of public goods will have a demonstration effect.
Network approach is necessary! municipalities exhibit interaction with neighbors
and may free-ride. Model interactions across municipalities as a network game,
Characterize its equilibria and identify structural parameters (Connects to
Blume’s talk in Kickoff and Task M1)
• These interactions imply that network effects are a major channels via which
state capacity affects economic development.
• Data-driven modeling, estimation, and theoretical analysis
Thrust 2:Task P2 Provision of development aid
• Field Experiment on development aid across 500 Afghan villages to look at the
effect of aid provision on security outcomes:
• Major Finding: public goods provision has a positive effect on solidifying
security in areas that are not highly insecure to begin with, but no effect in
highly insecure areas.
• Ongoing work: Effect of information campaigns on security outcomes.
Specifically, we are exploiting local random variation in radio signal to see
whether areas in our sample that receive both government services and
government radio are more secure.
• clean identification strategies for both public goods provision and for
government radio reception and can make causal arguments for our effects.
Task P3 Empirical Study of Social Interaction
Violence, Sectarianism and Communication in Yemen
Christia, Leskovec, Dahleh, Jadbabaie
Three years worth of call records data from Yemen (January 2010-January 2013; for
2-4 million subscribers; 5-10 million calls per day) to:
Assess the effects of drone strikes (exogenous violent shocks) and Arab Spring
protests (violent shocks endogenous to the society) on patterns of communication
(Christia, Leskovec)
Assess variation in patterns of communication during Shia, Sunni and joint religious
holidays in Yemen, to see how they vary across different geographic areas of the
country and how they compare against national celebrations, or other important
cultural or athletic events (Arab Idol, World Cup etc) (Christia, Dahleh, Jadbabaie)
Thrust 3: Micro-foundations: Strategic & Behavioral
models of Local Social Interactions
• Objective (Tasks M1-M3): Develop mathematical models of local interactions in
social, political and economic contexts and investigate implications of local
interactions on aggregate social and political outcomes.
• Key Questions discussed Last Year (Kleinberg, Leskovec, Kearns, Ozdaglar,
Jadbabaie, Jackson)
– Which patterns of local interactions enable spreading?
– What type of threshold model? Diminishing return vs critical mass?
– Competition and contagion, Financial cascades
This year: When is it possible to predict a cascade?
• Cascades are notoriously hard to predict.
• How well can we solve this prediction task
in a domain that is rich in features?
Cascades of photo re-shares on Facebook.
•
●
0.81
Mean Accuracy
•
•
●
●
Task: Given a cascade up to size k, predict if it will reach 2k.
Novel formulation as “growth prediction” leads
0.80
to balanced task.
Prediction accuracy up to 80%, increases as cascade 0.79
●
●
●
●
Total Reshares
● k or more
●
25
50
75
Number of reshares observed (k)
100
Thrust 3: Role of Network Structure on Diffusion:
• Empirical study of diffusion of microfinance in 75 rural villages in Karnataka,
India, relatively isolated from microfinance initially. BSS entered 43 of them and
offered microfinance
• Surveyed villages before entry, observed network structure and various
demographics. Tracked microfinance participation over time
• Villagers can identify highly central individuals, but how?
empirical validation of Jadbabaie/Jackson’s work in Task S1
•
•
Are people aware of network influence beyond their own
`neighborhood’? If so, how?
Answer this by showing that people will learn via gossip
about others’ centralities
T
DC( p,T) : ( pG) k
k 1
•

•
Agent i is initially informed,
each informed node tells
each of its neighbors with
prob p in each period, run for
T periods
Inverse problem: How do people learn other’s diffusion
centralities by observing local gossip?
Key insight: Gossip centrality and diffusion centrality
highly correlated. Connections to Thrusts 3,4
Thrust 3: Diffusion models:
• Task M1: New Models: Ozdaglar et al. Follow-up on Kleinberg: What type
of Threshold model is suitable for diffusion of innovations phenomena?
• Extended Relative/Absolute Stochastic Linear Threshold Mode
– What seed set maximizes overall diffusion ? (cf. Kearns 2013)
– What network properties make diffusion easier?
– Does history matter when agents make repeated adoption attempts?
– What differences are there between Relative and Absolute threshold
models?
• Empirical applications and testing using social network data
– ERSLTM: modeling mobile operator churn (with MIT Media Lab)
– EASLTM: adoption of a social app in overlapping online social networks
(with Humin Inc.)
Different threshold models, vastly different behaviors:
Data-driven, context-dependent modeling
– A-models: Information diffusion, supply-chain failure, social product adoption.
– R-models: Mobile operator churn, platform or technology adoption.
Thrust 4: Control of Sociopolitical Change
• Objective: Develop incentive and control mechanisms
• Networks provide a fundamental medium for the
propagation and diffusion of various forms of
contagion processes:
– e.g., opinion dynamics, epidemics, social cascades.
• We focus on dynamic control policies that use
information about the current state of the
contagion [Drakopoulos, Ozdaglar, Tsitsiklis 14].
• We combine tools from combinatorial optimization
and Markov Processes
 We provide a complete characterization of network
structures that are resilient against contagion
processes.
 We provide a new graph-theoretic measure called
“resistance” that captures how threatening a
subset of nodes is in terms of spreading the
contagion.
 We design efficient dynamic control policies
whenever possible.
Thrust 4: Sequential, networked global games
Motivation: self-fulfilling crises
• debt crises (PIGS)
• bank runs (Argentina 1999-2002)
• social upheavals (Arab revolutions)
• …
Connections to Thrust 1, Tasks S1,S2
Information sharing (locality) enables coordination.
How do equilibria depend on details of information sharing?
Dahleh 2013-2014, Shamma 2014
Network effects in global games
Dynamic effects
Steering Behavior via Rewards and Badges
Kleinberg-Leskovec
• How do we guide online (and
offline) behavior using rewards?
• Badges are a popular approach, but
optimal use of badges is not clear.
• A fundamental design question:
What kinds of behaviors can be
induced by badges?
Model of behavior in presence of
badges.
• Can understand process of
“acceleration” as a goal is
approached.
• Leads to a principled framework for
“badge placement” decisions:
optimal formulation of goals to
reward work.
•
•
•
•
•
•
•
•
•
Research personnel connections among
multiple institutions
Jadbabaie’s former undergrad Hallac working with Leskovec
at Stanford to analyze data obtained by Christia at MIT
Kearns’ former student doing a postdoc with Acemoglu &
Ozdaglar (Brautbar)
Jadbabaie’s former student (Molavi) planned to work with
Acemoglu over seummer as Econ PhD student
Jadbabaie and Kearns jointly advise student (Heidari)
Ozdaglar and Dahleh jointly advise student (Adams)
Acemoglu and Ozdaglar jointly advise Jadbabaie’s former
postdoc (Kakhbod)
Planned collaboration between Leskovec and Acemoglu
Leskovec collaboration with Kleinberg
Jadbabaie on Sabbatical at MIT, working with Acemoglu,
Ozdaglar, Dahleh, and Christia
Awards and honors
•
•
•
•
•
•
•
•
•
•
•
•
•
Daron Acemoglu received the Nemmers prize in economics
Daron Acemoglu has won the Kadir Has lifetime achievement award
Daron Acemoglu to receive an honorary PHD from Univ. of Athens
Daron Acemoglu received an Honorary doctorate from the Bosporus University
Ali Jadbabaie received the Alfred Fitler Moore Endowed Professorship in Network
Science from Penn’s School of Engineering and Applied Sciences
Ali Jadbabaie Elected as Inaugural Editor-in-Chief of the new IEEE Transactions on
Network Science and Engineering
Jon Kleinberg was elected as Simon’s Investigator
Asuman Ozdaglar received the inaugural Steven and Renee Finn innovation
fellowship from the MIT EECS Department
Jeff Shamma, received the UC Santa Barbara Mohammed Dahleh Award
&Distinguished Lecture
Matt Jackson gave the The Hahn Lecture, Royal Economic Society, April 2013;
Daron Acemoglu is the Editor-in-chief of Econometrica premiere journal in
Economic Sciences, Matt Jackon is the Co-editor
Jeff Shamma is the Editor at Large of new IEEE Transaction on Control of Network
Systems
Jure Leskovec received the 2013 WWW Best paper award
Education and Institutional Impact of the
MURI: Challenge and Opportunity
Interactions between engineered and natural physical systems, institutions,
and social & economic behavior as demonstrated by the research in this
MURI has created unprecedented challenges and opportunities magnified by
availability of large heterogeneous data on such interactions
Physical Systems
Social/Economic
Behavior
Institutions
21
A New research & educational entity at MIT
Human &
Institutional
Behavior
Information
and
Decision
Systems
Statistics
•
Energy
•
•
•
•
•
•
•
•
Transportation
Modeling & Prediction
Healthcare
Welfare & Sustainability
Financial Systems
Resilience & Systemic Risk
Social Networks
Design & Architecture
22
People and disciplines
LIDS New ESD
Hires
Human &
Institutional
Behavior
IDS
Statistics
Key MIT
Faculty
23
Education
Domain
knowledge
Questions
Complex systems
The Entity
21-Century
Statistics
Information and
Decision Sciences
Human and
Institutional
behavior
Academic Programs
Mathematical,
Behavioral and
Empirical
Science
24

similar documents