Causal model acquisition based on enumeration and pairwise interaction

Thumbnail Image



Li, Yu

Journal Title

Journal ISSN

Volume Title


University of Guelph


To construct a Bayesian Network, we need to acquire the conditional probability table (CPT) for each node. The complexity is exponential on the number of its parent nodes. The most widely used model to support such elicitation is the noisy-OR model. Then the complexity can be reduced to linear on the number of parent nodes. When multiple causes are present, they may either reinforce or undermine each other. The noisy-OR model can represent only reinforcement but not undermining. The recently proposed non-impeding noisy-AND tree model(NIN-AND)can represent both types of causal interactions, but it requires elicitation of the tree topology of causal interactions. When the number of causes is more than a few, it is mentally demanding to describe the topology accurately. This thesis presents two techniques to reduce the mental effort. One of them is based on tree enumeration and menu selection. The other is based on pairwise interaction between causes.



Bayesian Network, conditional probability table, noisy-OR model, non-impeding noisy-AND tree model, tree enumeration, menu selection, pairwise interaction