Recursive Classification of Multiple Objects Using Discordant and Non-Specific Data

The problem of multiple object classification based on discordant and non-specific data is considered. A general methodology for solving this problem is suggested and a suboptimal single-scan algorithm, referred to as the the global nearest neighbour (GNN), is implemented. The exact global dissimilarity measure, which is minimised by the GNN algorithm, is derived within the framework of the belief function theory. This measure, based on the plausibility of the global assignment, is related to the degree of conflict as understood in the transferable belief model interpretation of the belief function theory. The performance of the GNN algorithm was analysed by Monte Carlo simulations using different variants of the basic algorithm. One of the variants considered was the Bayesian GNN classifier. The results of this study suggest that the GNN classifier based on the exact global dissimilarity measure performs by far the best of the considered alternatives.