Approaches for Reducing Expert Burden in Bayesian Network Parameterization

Area of Science:

• Artificial Intelligence
• Probability and Statistics

Background:

• Bayesian networks (BNs) model complex variable relationships using conditional probability tables (CPTs).
• Structured Expert Judgment (SEJ) is used to elicit CPTs when data is insufficient, but it is often burdensome.
• Existing methods like InterBeta, Ranked Nodes Method (RNM), and Functional Interpolation aim to reduce this elicitation burden.

Purpose of the Study:

• To investigate the burden/accuracy trade-off of the InterBeta method for constructing CPTs.
• To compare InterBeta with RNM and Functional Interpolation.
• To propose and test extensions of the InterBeta method.

Main Methods:

• Reconstruction of previously elicited and simulated CPTs using InterBeta.
• Comparison of CPTs generated by InterBeta against those from RNM and Functional Interpolation.
• Testing of InterBeta extensions: shifted geometric mean, additional middle row elicitation, and the novel ExtraBeta extension.

Main Results:

• InterBeta with parent weights demonstrated superior performance in reconstructing CPTs.
• The ExtraBeta extension of InterBeta showed promising results for future research.
• The study evaluated the effectiveness of different methods in balancing elicitation burden and accuracy.

Conclusions:

• InterBeta, particularly with parent weights, is an effective method for reducing the burden of CPT elicitation in BNs.
• The proposed ExtraBeta extension warrants further investigation for its potential to improve CPT construction.
• Balancing accuracy and elicitation effort is crucial in expert judgment for probabilistic models.