An Adaptive k-NN Rule Based on Dempster-Shafer Theory

A new classiier using neighborhood information in the framework of the Dempster-Shafer theory of evidence has recently been introduced. This approach consists in considering each neighbor of a pattern to be classiied as an item of evidence supporting certain hypotheses concerning the class membership of that pattern. In this paper, an adap-tive version of this method is proposed, in which the parameters used to deene the basic probability assignments are learnt from the data by minimizing the mean squared error between the classiier outputs and target values. Based on the evidence-theoretic concepts of degree of connict and ignorance, new reject rules are introduced. Several sets of artiicial and real-world data are used for comparison with the voting and distance-weighted classiiers.