Extract Agent-based Model from Communication Network

Extract Agent-based
Model from
Communication
Network
Hung-Ching (Justin) Chen
Matthew Francisco
Malik Magdon-Ismail
Mark Goldberg
William Wallance
RPI
Goal
Given a society’s communication history,
can we:
n
Deduce something about “nature” of the society:
n
n
e.g., Do actors generally have a propensity to join
small groups or large groups?
Predict the society’s future:
n
e.g., How many social groups are there after 3
months?
n
e.g., What is the distribution of group size?
General Approach
Society’s
History
Society’s
Future
“Learn”
Individual
Behavior
(Micro-Laws)
“Predict”
(Simulate)
General Approach
Society’s
History
Society’s
Future
“Learn”
Individual
Behavior
(Micro-Laws)
“Predict”
(Simulate)
Social Networks
1
2
• Individuals
(Actors)
• Groups
3
Social Networks
1
2
• Individuals
(Actors)
- Join
- Leave
• Groups
3
Social Networks
1
• Individuals
(Actors)
4
- Join
- Leave
2
3
• Groups
- Disappear
- Appear
- Re-appear
Society’s History
General Approach
Society’s
History
Society’s
Future
“Learn”
Individual
Behavior
(Micro-Laws)
“Predict”
(Simulate)
Modeling of Dynamics
History
Parameters
Micro-Law
#1
Groups & Individuals
Micro-Law
#2
…
Actions
Micro-Law
#N
Join / Leave / Do Nothing
Example of Micro-Law
Actor X likes to join
SMALL
LARGE
Parameter
groups.
ViSAGE
Virtual Simulation and Analysis of Group Evolution
State
State: Properties of
Actors and Groups
State
State
update
Decide
Actors’ Action
Normative
Action
Actor
Choice
State
Feedback
to Actors
Real Action
Process
Actors’ Action
General Approach
Society’s
History
Society’s
Future
“Learn”
Individual
Behavior
(Micro-Laws)
“Predict”
(Simulate)
Learning
?
Parameters #1
in
Micro-Laws
?
Parameters #2
in
Micro-Laws
Learn
Communications
Groups & Group Evolution
Groups:
Overlapping
clustering
Communications
Group
evolution:
Matching
Groups
Evolution
Actor’s Types
n
Leader: prefer small group size and is most
ambitious
n
Socialite: prefer medium group size and is
medium ambitious
n
Follower: prefer large group size and is least
ambitious
Learning Actors’ Type
n
Maximum log-likelihood learning algorithm
n
Cluster algorithm
n
EM algorithm
Testing Simulation
Data
Testing Real Data
Learned Actors’ Types
Cluster
Algorithm
Leader
Socialite
Follower
Number of Actor
822
550
156
Percentage
53.8%
36.0%
10.2%
Learned Actors’ Types
EM
Algorithm
Leader
Socialite
Follower
Number of Actor
532
368
628
Percentage
34.8%
24.1%
41.1%
General Approach
Society’s
History
Society’s
Future
“Learn”
Individual
Behavior
(Micro-Laws)
“Predict”
(Simulate)
Testing & Simulations
Micro-Laws
&
Parameters
#1
Micro-Laws
&
Parameters
#2
Simulate
Simulate
Prediction
Prediction
Future Work
n
Test Other Predictions
n
n
e.g., membership in a particular group
Learn from Other Real Data
n
e.g., emails and blogs
Questions?