Quantiles based Neighborhood Method of Classification

Abstract

Classification of objects is an important problem that has received the attention of several researchers in Data Mining. Necessity for classification of an object into one of the predefined classes arises in several domains of research which include market research, document classification, diagnosing the presence of disease etc. A widely studied and applied popular classifying method which has attracted many data mining researchers is k-nearest neighbor algorithm. It is a distance based algorithm in which classification of an object is done on the basis of the memberships of its neighboring objects. The main problem one faces in the application classification is deciding a suitable value for the neighborhood parameter. In this paper, a method similar to classification in which the number of neighbors to be used in the classification process is determined by the distribution of distances between units in the training set has been proposed. Performance of the proposed method has been studied using simulated multivariate normal data sets as well as some benchmark data sets.