Probabilistic Graphical Models

Whether a Bayesian Network (BN) is constructed through expert elicitation, from data, or a combination of both, evaluation of the resultant BN is a crucial part of the knowledge engineering process. One kind of evaluation is to analyze how sensitive the network is to changes in inputs, a form of sensitivity analysis commonly called “sensitivity to findings”. The properties of d-separation can be used to determine whether or not evidence (or findings) about one variable may influence belief in a target variable, given the BN structure only. Once the network is parameterised, it is also possible to measure this influence, for example with mutual information or variance. Given such a metric of change, when evaluating a BN, it is common to rank nodes for either a maximum such effect or the average such effect. However this ranking tends to reflect the structural properties in the network: the longer the path from a node to the target node, the lower the influence, while the influence increases with the number of such paths. This raises the question: how useful is the ranking computed with the parameterised network, over and above what could be discerned from the structure alone? We propose a metric, Distance Weighted Influence, that ranks the influence of nodes based on the structure of the network alone. We show that not only does this ranking provide useful feedback on the structure in the early stages of the knowledge engineering process, after parameterisation the interest from an evaluation perspective is how much the ranking has changed. We illustrate the practical use of this on real-world networks from the literature.