Measures of conflicting evidence in Bayesian networks for classification

Classification selects one out of finitelymany classes based on available input and is applied to many kinds of application areas, e.g., diagnosis, monitoring, scoring, and pattern recognition. Often, classification is accomplished by use of Bayesian Networks which operate on evidence provided by sources, and calculate probabilities of classes as outcome. Sometimes, pieces of evidence from several sources provide substantially different but reliable information, indicating disparate classes. Such conflicting evidence has different reasons, e.g., measurement flaws, and can result in misclassification with severe consequences. Domain experts typically have a good intuitive understanding of evidence that should be judged as conflicting in given applications. Based on Bayesian Network literature, five measures for conflicting evidence are depicted and discussed. In an air defense scenario, these conflict measures are compared with domain experts' intuitive understanding of conflicts. Some aspects of configuration of such measures are discussed. Finally, coherence of different measures of conflicting evidence is surveyed in a second, maritime scenario.