Supervised cluster classification using the original n-dimensional space without transformation into lower dimension

A novel supervised classification algorithm, direct clustering in n-dimensi onal space (DCNS), was developed for difficult data sets where conventional methods of supervised clustering are expected to fail. The method is based , when applied on >3-dimensional spaces, on an algorithm that performs spec ial treatment on the measurement space, so that the treated space can allow a computer-aided clustering methodology similar to that used by human visi on, However, unlike other techniques that reduce the dimensionality of the space, the proposed method preserves the original dimensions while performi ng a computer-simulated human vision clustering in the original n-dimension al space. Thus the overlap between clusters that results from the dimension ality reduction is eliminated. The proposed method was applied to two real data sets. The results are compared with those obtained using principal com ponent analysis (PCA), an artificial neural network (ANN), and the k-neares t-neighbor (KNN) technique. On one data set containing only two clusters, t he DCNS algorithm gives better cluster separation than the other three meth ods. However, when all four methods were applied on the second data set, co ntaining eight different clusters, PCA, ANN and KNN were unable to give use ful cluster separation, while the DCNS method was able to separate all clus ters and classify the unknown points successfully with their corresponding clusters. The DCNS technique is able to perform other important cluster ana lysis tasks, such as testing the discriminatory power of a variable, select ing one variable from many, and conducting preliminary unsupervised cluster ing. Copyright (C) 2000 John Wiley & Sons, Ltd.