Cluster-based models of belief networks, social networks, and cultural knowledge Josh Tenenbaum, MIT 2007 MURI Annual Meeting Work of Charles Kemp, Chris Baker, Tom Griffiths, Pat Shafto, Vikash Mansinghka, Dan
47
Embed
Cluster-based models of belief networks, social networks, and cultural knowledge Josh Tenenbaum, MIT 2007 MURI Annual Meeting Work of Charles Kemp, Chris.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Slide 1
Cluster-based models of belief networks, social networks, and
cultural knowledge Josh Tenenbaum, MIT 2007 MURI Annual Meeting
Work of Charles Kemp, Chris Baker, Tom Griffiths, Pat Shafto,
Vikash Mansinghka, Dan Roy
Slide 2
Goal 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.
Slide 3
Approach Data Peoples beliefs about properties of objects
Relations between people Peoples beliefs about relations between
objects (or 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
Slide 4
Approach 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.
Slide 5
Talk 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 Application to Guatemala data
from Atran & Medin
Classic cluster-based methods Social networks: block
models
Slide 9
Classic cluster-based methods Cultural knowledge (joint
social/belief structure): cultural consensus model Not cluster-
based. SVD on matrix of people x questions
Slide 10
Problems 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 is often much
richer.
Slide 11
Talk 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 Application to Guatemala data
from Atran & Medin
Slide 12
people social relation Alyawarra tribe, central Australia
(Denham) 104 individuals 27 binary social relations 3 attributes:
kinship class, age, sex (used only for cluster validation, not
learning) people attributes Clustering arbitrary relational
systems
Slide 13
Infinite relational model (IRM) discovers 15 clusters
Clustering arbitrary relational systems
Slide 14
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, .
Slide 15
Slide 16
concept predicate Learning a medical ontology Data from UMLS
(McCrae et al.): 134 concepts: enzyme, hormone, organ, disease,
cell function... 49 predicates: affects(hormone, organ),
complicates(enzyme, cell function), treats(drug, disease),
diagnoses(procedure, disease)
Slide 17
Learning a medical ontology e.g., Diseases affect Organisms
Chemicals interact with Chemicals Chemicals cause Diseases
Slide 18
Hierarchical relational clustering
Slide 19
Slide 20
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.
Cross-cutting clustering with nested models
Slide 21
Conventional mixture model Cross-cutting clustering with nested
models
Slide 22
CrossCat model Cross-cutting clustering with nested models
Slide 23
Experimental tests of CrossCat = Stimuli: Task: Repeated free
sorting
Slide 24
Experimental tests of CrossCat Results: Conventional mixture
model CrossCat model Human frequency
Slide 25
Experimental tests of CrossCat Results: Conventional mixture
model CrossCat model Human frequency
Slide 26
Nested relational model: Cross-cutting clustering with nested
models people relation Infinite relational model: people relation
people relation
Slide 27
Talk 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 Application to Guatemala data
from Atran & Medin
Slide 28
Culture and cognition in Guatamela (Atran & Medin) Subjects
12 native Itza maya 12 immigrant Ladino 12 immigrant Qeqchi maya
Questions Does plant i help animal j? Does animal j help plant i?
animal plant 36 people 2 directions
Slide 29
Discovering cultural groups with the IRM animal plant 36 people
PA+
Slide 30
Cultural knowledge across groups animal plant 24 people (Itza,
Ladino) 2 directions
Slide 31
Ground Truth ecology
Slide 32
Cultural knowledge across groups Itza Ladino PA+AP+
Slide 33
I1 I2 I3 I5 I7 I8 I9 I10 I12 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 I4 Discovering
cultural groups with the nested IRM Data: PA+ Nesting structure
Cluster people Cluster plants within people Cluster animals within
plants and people Clusters of people found:
Discovering cultural groups with the nested IRM Data: AP+
Nesting structure Cluster people Cluster plants within people
Cluster animals within plants and people Clusters of people found:
L2 L3 L6 L7 L10 L11 L12 I1 I2 I3 I4 I5 I6 I7 I8 I9 I11 I12 L4 L5 L8
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12 L1 L9 I10
Slide 39
Conclusions 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
important qualitative structure in real-world data (Atran &
Medin).
Slide 40
Ongoing and future work More flexible nested structures More
dynamic data and analyses Second-generation Guatemala data
Political data sets: voting records, international relations More
structured representations necessary to capture cultural stories:
grammars, logical schemas (c.f. Forbus, Richards, Atran) people
plantsanimals directionality
a(3). a(9). a(1). a(13). a(5). a(11). b(7). b(14). b(2). b(10).
b(6). c(12). c(4). c(8). c(15). r(X,Y) a(X),a(Y). (0.0) r(X,Y)
a(X),b(Y). (0.9) r(X,Y) c(X),a(Y). (0.95)... r(3,7). r(1,10).
r(2,4)... The concepts discovered by the IRM can serve as
primitives in complex logical theories (cf. clustering approaches
to predicate invention, e.g., Craven and Slattery (2001) or
Popescul and Ungar (2004)) 3 9 1 13 5 11 7 14 2 10 6 12 4 8 15 3 9
1 13 5 11 7 14 2 10 6 12 4 8 15
Slide 47
Related Work Relational models Sociology: Wang and Wong (1987);
Nowicki and Snijders (2001) Machine learning: Taskar, Segal and
Koller (2001) Wolfe and Jensen (2004) Wang, Mohanty and McCallum
(2005) Nonparametric Bayesian models Ferguson (1973); Neal (1991)
Nonparametric Bayesian relational models Carbonetto, Kisynski, de
Freitas and Poole (2005) Xu, Tresp, Yu, Kriegel (2006)
Slide 48
Optimization (or inference) Global proposals Split and merge
clusters Local proposals Re-assign one entity to best cluster based
on current assignments of all other entities (i.e., Gibbs sampling)
Both cognitively plausible and computationally reasonable.
Independent symmetric beta priors on : Chinese Restaurant
Process over z: Goal: Infer Infer (integrating out to reduce space
of unknowns) Generating and z
Slide 53
Global-local search process
Slide 54
Joint modeling of belief systems and social systems animal
plant person helps(plant,animal,person judging) Data from Atran and
Medin