Elicitation Techniques for Bayesian Network Models

The document describes two techniques for creating elicited probability tables (EPTs) to simplify the process of building Bayesian networks when measured data is unavailable or insufficient. Both methods utilize a user-defined "typical distribution" to represent a normal state of affairs. The first method, "Influence Weights," assumes that changes in parent nodes independently affect the child node and is best used when parent and child nodes share the same states. The second method, "Likelihood," uses a log-likelihood approach to model how parent nodes shift the child node's probability away from the typical distribution, making it suitable for cases where parent nodes have different states than the child node.

• Expert elicitation of CPTs is necessary when there is no measured data to estimate the table or when the data set is too small relative to the number of parameters to definitively establish the CPT.
• The Influence Weights method assumes that the change in probability resulting from a change in any parent node is independent of other parent nodes, meaning a change from "low" to "high" in one parent produces the same effect regardless of other parent states.
• To account for influences other than identified parent nodes in the Influence Weights method, an "environmental" parent node is added that always maintains the typical distribution.
• The Influence Weights elicitation process involves three steps: identifying the most influential parent node (assigned a non-normalized weight of 10), assigning relative weights to other nodes, and estimating the total "information" (I) as a percentage of certainty about the child node.
• The Likelihood method is based on Bayes' rule and assumes that the log likelihood can be expressed as a sum of independent terms for each parent node, shifting the child node's probability away from the typical distribution.
• The Influence Weights method is most appropriate when parent nodes share the same states as the typical distribution and the expert believes typical inputs should yield a typical output.
• The Likelihood method is preferred when the expert's mental model is that parent nodes drive the child node away from a typical distribution, such as rainfall levels shifting the probability distribution of crop yield.

APA

Kemp-Benedict, E. (2008). Elicitation Techniques for Bayesian Network Models. Stockholm Environment Institute. https://www.sei.org/mediamanager/documents/Publications/Climate/elicitation-techniques-bayesian.pdf

Chicago

Kemp-Benedict, Eric. Elicitation Techniques for Bayesian Network Models. Stockholm Environment Institute, 2008. https://www.sei.org/mediamanager/documents/Publications/Climate/elicitation-techniques-bayesian.pdf.

Wikipedia

{{cite report |last1=Kemp-Benedict |first1=Eric |title=Elicitation Techniques for Bayesian Network Models |publisher=Stockholm Environment Institute |date=6 June 2008 |url=https://www.sei.org/mediamanager/documents/Publications/Climate/elicitation-techniques-bayesian.pdf |access-date=23 October 2026 |via=Climate Insights Directory}}

BibTeX

@techreport{kempbenedict2008elicitation, author = {Kemp-Benedict, Eric}, title = {{Elicitation Techniques for Bayesian Network Models}}, institution = {Stockholm Environment Institute}, year = {2008}, month = jun, url = {https://www.sei.org/mediamanager/documents/Publications/Climate/elicitation-techniques-bayesian.pdf}, urldate = {2026-10-23}, note = {Indexed by Climate Insights Directory} }