NEAREST NEIGHBOR CLASSIFIERS WITH IMPROVED ACCURACY AND EFFICIENCY

Abstract

Nearest Neighbor algorithms are non-parametric algorithms that use distance measure techniques for classification and regressions. This thesis uses the method of pruning to improve accuracy and efficiency of a nearest neighbor classifier and also states the different stages the pruning algorithm can be applied and shows the best stage for pruning which gives the maximum accuracy. The performance of the classifier is shown to be better than other improved nearest neighbor classifiers. A fast method of finding the optimal k in a k-nearest neighbor classifier is proposed in the thesis. A method of optimizing the distance measure using a second order training algorithm in a k-nearest neighbor algorithm is also proposed in this thesis which results to better accuracy than the traditional k-nearest neighbor classifier.