Title: Infinite block models for belief networks, social networks, and cultural knowledge
1Infinite block models for belief networks, social
networks, and cultural knowledge
- Josh Tenenbaum, MIT
- 2007 MURI Review Meeting
- Work of Charles Kemp, Chris Baker, Tom Griffiths,
Pat Shafto, Vikash Mansinghka, Dan Roy
2Goal
- Algorithmic tools for uncovering structure in
belief networks, social networks, and joint
structure (social-belief networks). - Why?
- Joint social-belief structure culture
- Algorithms let us map cultural knowledge quickly
and semi-automatically, detect changes and track
dynamics.
3Approach
- Data measurement matrices
- Peoples beliefs about things in their
environment - Relations between people
- Peoples beliefs about relations between things,
or relations between people. - Representation cluster-based models
- Clusters of things categories
- Clusters of people social groups
- Clusters of people who share similar beliefs
about clusters of things or people cultural
groups
4Approach
- Learning Bayesian inference from data
- Relational models analyze arbitrary relational
databases of beliefs, not just a single matrix - Nonparametric models automatically determine
complexity of representations - Hierarchical models multiple levels of structure
- Nested models structures with structure
- Result a flexible toolkit that goes
qualitatively beyond standard algorithms. - e.g., ability to discover cultural groups
characterized by a shared understanding of the
environments relational structure.
5Talk outline
- Classic cluster-based methods
- New methods
- Clustering with arbitrary relational systems
- Hierarchical relational clustering
- Cross-cutting clustering with nested models
- Cross-cutting relational clustering
- Guatemalan cultural data from Atran Medin
- Conclusions and future directions
6Classic cluster-based methods
- Belief networks mixture models
7Classic cluster-based methods
- Belief networks mixture models
8Classic cluster-based methods
- Social networks block models
DefersTo(Pi, Pj)
9Classic cluster-based methods
- Cultural knowledge (joint social/belief
structure) cultural consensus model
Not cluster-based. SVD on matrix of people x
questions
10Problems with classic methods
- No principled tools for discovering different
cultural groups characterized by different belief
networks. - CCM not intended to find cultural groups, but
rather to uncover (and test for) shared knowledge
and authoritativeness in a single cultural group.
Test theory without an answer key - Can only represent simple forms of knowledge that
fit into a single two-mode matrix. - Cultural knowledge and social contexts are often
much richer.
11Talk outline
- Classic cluster-based methods
- New methods
- Clustering with arbitrary relational systems
- Hierarchical relational clustering
- Cross-cutting clustering with nested models
- Cross-cutting relational clustering
- Guatemalan cultural data from Atran Medin
- Conclusions and future directions
12Clustering arbitrary relational systems
people
people
social relation
people
attributes
- Alyawarra tribe, central Australia (Denham)
- 104 individuals
- 27 binary social relations
- 3 attributes kinship class, age, sex (used
only for cluster validation, not learning)
13Clustering arbitrary relational systems
Infinite relational model (IRM) discovers 15
clusters
14Clustering arbitrary relational systems
- International relations circa 1965 (Rummel)
- 14 countries UK, USA, USSR, China, .
- 54 binary relations representing interactions
between countries exports to( USA, UK ),
protests( USA, USSR ), . - 90 (dynamic) country features purges, protests,
unemployment, communists, languages,
assassinations, .
15(No Transcript)
16Clustering arbitrary relational systems
concept
predicate
concept
- Data from UMLS (McCrae et al.) medical knowledge
base - 134 terms enzyme, hormone, organ, disease, cell
function ... - 49 predicates affects(hormone, organ),
complicates(enzyme, cell function), treats(drug,
disease), diagnoses(procedure, disease)
17Clustering arbitrary relational systems
Conceptual clusters discovered
18Concept maps
19Hierarchical relational clustering
20Hierarchical relational clustering
21Cross-cutting clustering with nested models
- Models so far all learn a single system of
clusters. - We would like to be able to discover multiple
cross-cutting systems of clusters. - Within an individuals mind multiple mental
models of a single complex domain. - Across individuals in a population multiple
cultural groups with different characteristic
mental models.
22Cross-cutting clustering with nested models
Conventional mixture model
23Cross-cutting clustering with nested models
CrossCat model
24Analysis of US Senate votes 1989-90
- 101 senators x 638 issues 10 systems of
classes.
Environment agriculture
Hot-button social issues
Core democratic platform
Law and order
Military
25Cross-cutting clustering with nested models
Nested relational model
Infinite relational model
people
people
relation
relation
people
people
26Discovering cultural groups based on shared
relational knowledge
- Guatemala studies of Atran Medin
- Subjects
- 12 native Itza maya
- 12 immigrant Ladino
- 12 immigrant Qeqchi maya
- Questions
- Does plant i help animal j?
Nested relational model
animal
people
plant
27Discovering cultural groups based on shared
relational knowledge
Clusters of people found
- Guatemala studies of Atran Medin
- Subjects
- 12 native Itza maya
- 12 immigrant Ladino
- 12 immigrant Qeqchi maya
- Questions
- Does plant i help animal j?
L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11 L12 I6 I11
Q3 Q6 Q8 Q9 Q10 Q11 Q12
Q1 Q2 Q4 Q5 Q7
I1 I2 I3 I5 I7 I8 I9 I10 I12
I4
28Talk outline
- Classic cluster-based methods
- New methods
- Clustering with arbitrary relational systems
- Hierarchical relational clustering
- Cross-cutting clustering with nested models
- Cross-cutting relational clustering
- Guatemalan cultural data from Atran Medin
- Conclusions and future directions
29Conclusions
- A flexible toolkit for statistical learning about
cultural knowledge and cultural groups. - Relational models analyze arbitrary relational
databases of beliefs, not just a single matrix - Nonparametric models automatically determine
complexity of representations - Hierarchical models multiple levels of structure
- Nested models structures with structure
- Can automatically discover cultural structure in
real-world data (Atran Medin, DARPA CPoF).
30Ongoing and future work
- Algorithms that can scale to very large networks.
- More dynamic data and models.
- Second-generation Guatemala data
- Political data sets voting records,
international relations - Better statistical models for sparse networks.
- Integrated models of values and beliefs.
- More structured representations necessary to
capture cultural stories grammars, logical
schemas. - Multi-level statistical models for learning about
network structure from raw event data, or
learning about networks with different forms of
structure.
31Learning network structure from raw event data
of samples 20 80
1000
Network N
edge (N)
Data D
Classes Z
1 2 3 4 5 6
7 8 9 10 11 12 13 14 15 16
class (Z)
Abstract Classes
c1
c2
c1
c2
h
0.4
0.0
c1
0.0
0.0
c2
edge (N)
Network N
Data D
(Mansinghka, Kemp, Tenenbaum, Griffiths UAI 06)
32Learning network structure from raw event data
12
1
11
2
10
3
9
4
8
5
7
6
of samples 40 100
1000
Network N
edge (N)
Data D
Classes Z
c1
h
class (Z)
1 2 3 4 5 6 7 8 9 10 11 12
Abstract Classes
0.1
c1
c1
edge (N)
Network N
Data D
(Mansinghka, Kemp, Tenenbaum, Griffiths UAI 06)
33Learning different forms of network structure
Dominance hierarchy Tree
Cliques Ring
Primate troop Bush administration
Prison inmates New Guinea islands
beats told likes
trades with