Possibilistic network-based classifiers: On the reject option and concept drift issues

Abstract

In this paper, we deal with two important issues regarding possibilistic network-based classifiers. The first issue addresses the reject option in possibilistic network-based classifiers. We first focus on simple threshold-based reject rules and provide interpretations for the ambiguity and distance reject then introduce a third reject kind named incompleteness reject occurring when the inputs are missing or incomplete. The second important issue we address is the one of concept drift. More specifically, we propose an efficient solution for revising a possibilistic network classifier with new information. © 2011 Springer-Verlag.