A Multiple Neural Network Architecture for Sequential Evidence Aggregation and Incomplete Data Classification

In this paper, a multiple neural network architecture is proposed for undertaking the problems associated with incomplete or missing data in on-line learning and classification tasks. An autonomously learning neural network classifier, which has been previously devised based upon integration of Fuzzy ARTMAP and the Probabilistic Neural Network, is employed as the basis for the development of the multiple neural network system. Each classifier is dedicated to handle a set of input features independently and produces a prediction of the target class. Bayes' theorem is then applied to combine the outcomes from disparate classifier modules sequentially. Applicability of the multiple neural network system is demonstrated using a simulated data set and a real medical diagnosis database and the results are compared with other approaches.