Test-selection strategies to increase the number of MAP-independent variables
In practical applications, decision support systems that motivate and justify decisions are essential for a user to understand and accept the systems. Recently, the concept of MAP-independence was introduced which aimed to provide such justification for Bayesian networks by explicating the relevant intermediate variables for decisions in MAP problems. Whereas the concept of MAP-independence may work well to justify a decision when the number of relevant variables is small, in this paper we argue that for scenarios in which the number of relevant variables is large, these justifications would be inadequate and a user can only gain limited insight into the decision. Therefore, we focus on finding test-selection strategies that can decrease the number relevant variables in the network by deciding which variable to gather evidence for in order to obtain a decision that is better to justify and motivate. Specific test-selection strategies that we investigate include the in-degree, out-degree and total degree of a variable, the distance of a variable to the explanation variable, the expected utility, the expected Gini index and mutual information with the explanation variable. After running systematic simulations on the ALARM network and analysing the results based on rank- and value-approximation, we conclude that the distance to the explanation variable, expected utility, expected Gini index and mutual information could serve as good test-selection strategies to decrease the number of relevant variables. However, more research is needed to be able to generalize these findings to a larger population of Bayesian networks.
Faculteit der Sociale Wetenschappen