Visualizing Inference in Bayesian Networks

Inference in Bayesian networks is used to calculate the posterior probability distributions of unobserved variables in a network. These posterior probability distributions are used to draw conclusions and are the basis for decisions, in the domain of a particular model. Inference is a complex process and can be difficult to understand for even the most experienced Bayesian network users. In this thesis, we propose a technique to visualize important aspects of a Bayesian network, in order to make the process of inference more insightful. This technique consists of augmenting the visual representation of a Bayesian network with extra information. The only function of arcs in a Bayesian network is to indicate the relationships among the variables. We have used the arcs in a Bayesian network to show additional information: (1) the thickness of an arc is automatically adjusted to represent the strength of influence between two directly connected nodes and (2) the color of an arc is automatically adjusted to indicate the sign of influence between two directly connected nodes. Our technique does this in a novel, dynamic way, which is context-specific and takes into account any indirect influences. We have implemented our technique and integrated it into a software package called GeNIe, which can be used for developing Bayesian networks and is developed at the Decision Systems Laboratory of the University of Pittsburgh. A qualitative empirical evaluation showed that our technique and implementation are easy to use and understand and give a user more insight into a particular Bayesian network.