Reduction of clustering problem to pattern recognition

Abstract The computing time required by most of the clustering programs becomes prohibitively long as the number of objects to be classified increases. It is shown to be effective in overcoming this difficulty to select a small number of “representative” objects first and to apply the clustering program on them. The non-representative objects are thereafter placed in the generated classes by the pattern recognition technique, where the role of paradigms (class-samples) is played by the representative objects. The representative objects are those which have large components in the feature-subspace in the sense of SELFIC.