THEORETICAL-ANALYSIS OF THE CONFIDENCE METRICS FOR NEAREST-NEIGHBOR CLASSIFIER

Confidence value plays a vital role in the decision of rejection thres hold and the integration of multiple classifiers. Nearest neighbor (NN ) classifier is the most traditional and most common nonparameter stat istical pattern classifier, However, so far there is no explicate theo retical analysis of the connection between nearest distance and confid ence value. An analytical insight into different approximations is pre sented and one formula is pointed out that it is optimal in the sense of mathematical expectation. Practice in handwritten numeral recogniti on strongly supports the conclusion.