Two new nearest neighbor classification rules

Abstract

Nearest Neighbor (NN) classification is a non-parametric discrimination and classification technique. In NN classification a test item is compared by some similarity measure of its multiple variables (usually a distance metric) with all the items in a training set. The class of the item to which it is most similar can be used as an indication of the class of the test item. In other words, the test item is assigned the class of its nearest neighbor. A key extension is the case when k > 1 nearest neighbors (k-NN) are examined with the classification usually being made based on a plurality. NN classification is used in many fields, including for example the field of Pattern Recognition. Applications include tasks like speech recognition by a computer, medical data interpretation and diagnosis, or the interpretation of remote sensing imagery from satellites. Military applications of the technique include any situation were automated recognition is required. This thesis proposes two new NN rules that are intended to improve classification accuracy. The rules are tested against baseline classification methods in common use with a variety of data sets. One method shows improvement over the baseline methods in most of the data cases examined.