Improving 1NN strategy for classification of some prokaryotic organisms

Abstract

Classification algorithms are intensively used in discovering new information in large sets of biological data. In cases when classification tasks involve nominal attributes, some of commonly used classification tools do not obtain results of satisfying quality, since mathematical operations and relations can not be directly applied to symbolic values. This problem often appears in the k-nearest neighborhood (KNN) classification because the standard Euclidean distance function can become burdened by the large number of irrelevant attributes, consequently producing inaccurate classification results. In this paper we examine several metrics which can be applied to nominal attributes and for each metric we apply the appropriate KNN strategy. In order to justify the proposed approach, comprehensive experiments are performed on a dataset of prokaryiotic organisms. Experimental results indicate that the new classifications are more accurate than those obtained by the previously used methods, getting better results in seven of total of twelve cases.