Using Dempster-Shafer Theory in MCF Systems to Reject Samples

In this paper the Dempster-Shafer theory of evidence is utilised in multiple classifier systems to define rejection criteria for samples presented for classification. The DS theory offers the possibility to derive a measure of contradiction between the classifier decisions to be fused. Moreover, assigning positive belief mass to the universal hypothesis Θ in the basic probability assignments produced by the classifiers, allows to quantify the belief in their correctness. Both criteria have been evaluated by numerical simulations on two different benchmark data sets. The results are compared to standard static classifier combination schemes and basic classifiers. It is shown that DS classifier fusion can boost the combined classifier accuracy to 100% on the set of accepted data points (~ 70%). This behaviour could be of interest in applications with high costs for a miss, e.g. in medical screening tests.