Evidence aggregation for diagnosis: Bayesian and fuzzy strategies

This paper investigates the utility of using aggregate (average, max etc.) information in a multi-sample classification problem in terms of the accuracy of the overall classification, and the prediction of estimated error. Bayesian networks are presented here in order to allow comparison of these results with those of a previously presented fuzzy inference system. Different structures of Bayesian network are implemented, and also compared. Different methods of calculating aggregate information to append to the samples are also investigated. The authors find that appending the additional information does improve classification accuracy and lowers classification frequency. They also find that the addition of the average aggregator to the data is more effective than appending any other combination of aggregators. Lastly, they find that the naive Bayesian network structure performs better than tree-augmented Bayesian networks, as well as networks limited only in number of parents per node.