Optimal Node Ordering in Bayesian Belief Networks

Eliciting probability distributions from experts has always been one of the most difficult challenges in building Bayesian Belief Networks. Most of the existing literature has focused on different tools to use during the elicitation process, or various cognitive biases to avoid. However, there are very few guidelines on the order in which nodes that should be presented to the experts during the elicitation process, or how experts react to different orders of nodes. There is also a lack of research on the potential use of Internet webpage as a new means for the probability elicitation process.

In this tutorial paper, we developed some models for determining the order in which the nodes are to be elicited based on various assumptions. We conducted a probability elicitation experiment with some graduate students using an online elicitation tool that we have developed in order to explore the implications of our node-ordering models and the potential of eliciting probability through an online webpage.

The experimental results indicate that 41% of the students experienced fatigue even during a 10-minute online probability elicitation survey, and that the node ordering in which the number of conditioning parents is increasing is the user-friendliest model out of the three different node-ordering models that we have created. It may also be possible to mitigate the difficulty of probability elicitation that arises from the high number of conditioning parents by changing the node ordering. These experimental results show that the selection of node ordering for the probability elicitation process cannot be taken for granted given a large Bayesian Belief Network, and further research on this subject is worth exploring.