## Discovering latent classes in relational data (2004)

Citations: | 40 - 4 self |

### BibTeX

@TECHREPORT{Kemp04discoveringlatent,

author = {Charles Kemp and Thomas L. Griffiths and Joshua B. Tenenbaum},

title = {Discovering latent classes in relational data},

institution = {},

year = {2004}

}

### Years of Citing Articles

### OpenURL

### Abstract

We present a framework for learning abstract relational knowledge with the aim of explaining how people acquire intuitive theories of physical, biological, or social systems. Our approach is based on a generative relational model with latent classes, and simultaneously determines the kinds of entities that exist in a domain, the number of these latent classes, and the relations between classes that are possible or likely. This model goes beyond previous psychological models of category learning, which consider attributes associated with individual categories but not relationships between categories. We apply this domain-general framework to two specific problems: learning the structure of kinship systems and learning causal theories. 1 1

### Citations

1991 |
Social Network Analysis: Methods and Applications
- Wasserman, Faust
- 1994
(Show Context)
Citation Context ...cognitive relevance of other relational structures discussed in these fields, such as the overlapping class model of Kubica et al. [16], and structures where groups are defined by regular equivalence =-=[17]-=-. Developing a framework that can form spontaneous and flexible combinations of these structures remains a formidable open task. Acknowledgments We thank Woodrow Denham for supplying the Alyawarra dat... |

456 | Attention, similarity, and the identification-catgorization relationship
- Nosofsky
- 1986
(Show Context)
Citation Context ...— problems involving attributes alone, relations alone, or some combination of the two. Psychologists have paid most attention to the first problem, and have developed successful computational models =-=[1, 2]-=- that explain how people discover classes given attribute data. Our approach handles all three problems, although it reduces to Anderson’s rational model of categorization [1] when given only attribut... |

233 | The adaptive nature of human categorization
- Anderson
- 1991
(Show Context)
Citation Context ...— problems involving attributes alone, relations alone, or some combination of the two. Psychologists have paid most attention to the first problem, and have developed successful computational models =-=[1, 2]-=- that explain how people discover classes given attribute data. Our approach handles all three problems, although it reduces to Anderson’s rational model of categorization [1] when given only attribut... |

163 | The infinite Gaussian mixture model
- Rasmussen
- 2000
(Show Context)
Citation Context ... the model in which z is generated according to Equation 2 and G is generated according to Equation 1 the infinite blockmodel. Other infinite models have been proposed by machine learning researchers =-=[5, 6]-=- using a similar construction. 2.3 Model inference Having defined a generative model for G and z, we use Bayesian inference to compute a posterior distribution over z given G: (2) p(z|G) ∝ p(G|z)p(z) ... |

129 |
Estimation and prediction for stochastic blockstructures
- Nowicki, Snijders
- 2001
(Show Context)
Citation Context ...olds between a pair of objects with some probability that depends only on the classes of those objects. Statisticians and sociologists have used a model of this kind, called the stochastic blockmodel =-=[3, 4]-=-, to analyze social networks. The stochastic blockmodel, however, assumes a fixed, finite number of classes. When discovering the latent structure of a domain, people learn the number of classes at th... |

109 | Probabilistic classification and clustering in relational data
- Taskar, Segal, et al.
- 2001
(Show Context)
Citation Context ... Our framework for discovering latent classes is an extension of relational models previously proposed in sociology (stochastic block models [4]) and machine learning (probabilistic relational models =-=[15]-=-). We are also exploring the cognitive relevance of other relational structures discussed in these fields, such as the overlapping class model of Kubica et al. [16], and structures where groups are de... |

99 | A split-merge Markov chain Monte Carlo procedure for the Dirichlet process mixture model
- Jain, Neal
(Show Context)
Citation Context ...uation 2. To facilitate mixing, we supplement our Gibbs sampler with two Metropolis-Hastings updates. First, we consider proposals that attempt to split a class into two or merge two existing classes =-=[8]-=-. Split-merge proposals allow sudden large-scale changes to the current state rather than the incremental changes characteristic of Gibbs sampling. Second, we run a Metropolis-coupled Markov chain Mon... |

61 |
The Elementary Structures of Kinship
- Levi-Strauss
- 1969
(Show Context)
Citation Context ... structure of their kinship systems. In a mathematical appendix to a work by Levi-Strauss, André Weil notes that several Australian kinship systems are isomorphic to the dihedral group of order eight =-=[9]-=-. Even trained field workers find these systems difficult to 7s1 3 mother father father mother Figure 4: Relations between the four sections of a Kariera kinship system. Members of section 4 have moth... |

55 | Stochastic link and group detection
- Kubica, Moore, et al.
- 2002
(Show Context)
Citation Context ...ng (probabilistic relational models [15]). We are also exploring the cognitive relevance of other relational structures discussed in these fields, such as the overlapping class model of Kubica et al. =-=[16]-=-, and structures where groups are defined by regular equivalence [17]. Developing a framework that can form spontaneous and flexible combinations of these structures remains a formidable open task. Ac... |

33 | Bayesian mixture modeling by Monte Carlo simulation
- Neal
- 1991
(Show Context)
Citation Context ... the model in which z is generated according to Equation 2 and G is generated according to Equation 1 the infinite blockmodel. Other infinite models have been proposed by machine learning researchers =-=[5, 6]-=- using a similar construction. 2.3 Model inference Having defined a generative model for G and z, we use Bayesian inference to compute a posterior distribution over z given G: (2) p(z|G) ∝ p(G|z)p(z) ... |

21 | Learning causal laws
- Tenenbaum, Niyogi
- 2003
(Show Context)
Citation Context ...bjects could discover these simple theories after interacting with some subset of the objects. Their experiments had seven phases and three new objects were added to the screen during each phase (see =-=[14]-=- for details). As new objects were added, subjects made predictions about how these objects would interact with old objects or with each other. At the end of the 11sa 3 b 4 a 4 ... ? a 3 ? a 4 ... ? b... |

16 |
Stochastic a posteriori blockmodels: Construction and assessment
- Wasserman, Anderson
- 1987
(Show Context)
Citation Context ...olds between a pair of objects with some probability that depends only on the classes of those objects. Statisticians and sociologists have used a model of this kind, called the stochastic blockmodel =-=[3, 4]-=-, to analyze social networks. The stochastic blockmodel, however, assumes a fixed, finite number of classes. When discovering the latent structure of a domain, people learn the number of classes at th... |

12 |
Aranda and Alyawarra Kinship: A Quantitative Argument for a Double Helix Model. American Ethnologist
- Denham, McDaniel, et al.
- 1979
(Show Context)
Citation Context ... appear to have no explicit representations of kinship rules, let alone cultural transmission of such rules. 2 We focus here on the Alyawarra, a Central Australian tribe studied extensively by Denham =-=[11, 12, 13]-=-. Using Denham’s data 3 , we show that our model is able to discover some of the properties of the Alyawarra kinship system. The Alyawarra have four kinship sections and the Alyawarra language contain... |

10 |
The detection of patterns in Alyawarra nonverbal behavior
- Denham
- 1973
(Show Context)
Citation Context ... appear to have no explicit representations of kinship rules, let alone cultural transmission of such rules. 2 We focus here on the Alyawarra, a Central Australian tribe studied extensively by Denham =-=[11, 12, 13]-=-. Using Denham’s data 3 , we show that our model is able to discover some of the properties of the Alyawarra kinship system. The Alyawarra have four kinship sections and the Alyawarra language contain... |

3 |
Multiple Measures of Alyawarra Kinship, Field Methods 17(1):70-101. Guest edited by Dwight Read. See further analysis of these data by members of the machine learning team at the MIT Department of Brain and Cognitive Sciences
- Denham, White
- 2004
(Show Context)
Citation Context ... appear to have no explicit representations of kinship rules, let alone cultural transmission of such rules. 2 We focus here on the Alyawarra, a Central Australian tribe studied extensively by Denham =-=[11, 12, 13]-=-. Using Denham’s data 3 , we show that our model is able to discover some of the properties of the Alyawarra kinship system. The Alyawarra have four kinship sections and the Alyawarra language contain... |

2 | Automatic rule discovery for field work in anthropology
- Findler
- 1992
(Show Context)
Citation Context ...r sections of a Kariera kinship system. Members of section 4 have mothers in section 1 and fathers in section 2. This simple model is consistent with the relational data shown in Figure 5. understand =-=[10]-=- which raises an intriguing question about cognitive development: how do children discover the social structure of their tribe? The learning problem is particularly interesting since some communities ... |

1 | Alyawarra ethnographic database: Numerical data documentation file, version 7. Alyawarra - Denham - 2001 |