Tutorial on How to Measure Link Strengths in Discrete Bayesian Networks

This report discusses measures for link strength in Discrete Bayesian Networks, i.e. measures for the strength of connection along a specific edge. It is a revised version of Report GT-IIC-07-01 (Jan 2007) with improved literature review and explanations. The target application is the visualization of the strengths of the edge connections in a Bayesian Network learned from data to learn more about the inherent properties of the system. The report reviews existing link strength measures, provides an accessible derivation of the primary measure, proposes some simple variations of the primary measure and compares their resulting properties. 1 The Concept of Link Strength Boerlage was the first to formally introduce the concept of link strength for Bayesian Networks (Boerlage 1992). Boerlage defines connection strength for any pair of nodes (adjacent or not) to measure the strength between those nodes taking any possible path between them into account. In contrast link strength (also known as arc weight) is defined for a specific edge and measures the strength of connection only along that single edge. To demonstrate the difference between these concepts in particular for adjacent nodes consider the network in Figure 1. Each of the three nodes only has two states, True and False. Let us focus on the connection Joint appointment with the Robotics and Intelligent Machines Center, School of Interactive Computing, College of Computing, Atlanta, Georgia 30308. X