Constructing a Bayesian Network requires the conditional probabilities table (CPT)
to be acquired, one for each variable or node in the network. When data mining is not
available, CPTs must be acquired from the domain experts. The complexity of the
direct elicitation is exponential on the number of parents of a variable, making direct
elicitation from human experts impractical for a large number of causes. Causal models
such as Noisy-OR, Noisy-AND, Noisy-MIN, Noisy-MAX and Recursive Noisy-OR
have been developed that allow CPTs acquisition to be achieved with linear complexity
on the number of causes. Their representation power is measured by their ability
to encode the causal interactions. Causal interactions can be categorized into two
types: reinforcing and undermining. The Non-Impeding Noisy-AND or NIN-AND
tree causal model, developed by Xiang and Jia, is capable of modeling both types of
interaction while retaining the linear complexity. The main challenge in utilizing the
NIN-AND tree model to generate a CPT is that it requires its tree topology to be
elicited. A NIN-AND tree topology is an encoding of the causal interactions between
the causes. In this work we present two methods, Structure Elimination (SE) and
Pairwise Causal Interaction (PCI), that allow indirect elicitations of the NIN-AND
tree topology using some additional probabilities elicited from experts. We conduct
human-based experiment to investigate the e ectiveness of the two methods in terms
of accuracy by comparing them to the Direct Numerical (DN) elicitation method. We
recruit participants from second year Computer Science students at the University
of Guelph. The process involves training a participant into domain expert using a
known NIN-AND tree model then acquire another NIN-AND tree model by applying
the SE and PCI methods. The CPTs produced by the acquired NIN-AND tree models
are then compared to the one obtained by using the DN method. Comparable CPT
accuracies are obtained among models generated by di erent methods, even though
SE and PCI requires a much smaller number of parameters in comparison to DN.
Identifer | oai:union.ndltd.org:LACETR/oai:collectionscanada.gc.ca:OGU.10214/3011 |
Date | 15 September 2011 |
Creators | Truong, Minh |
Contributors | Xiang, Yang |
Source Sets | Library and Archives Canada ETDs Repository / Centre d'archives des thèses électroniques de Bibliothèque et Archives Canada |
Language | English |
Detected Language | English |
Type | Thesis |
Page generated in 0.0016 seconds