Sequentially grouping items into clusters of unspecified number

Abstract

When run, most traditional clustering algorithms require the number of clusters sought to be specified beforehand, and all clustered items to be present. These two, for practical applications very serious shortcomings are overcome by a straightforward sequential clustering algorithm. Its most crucial constituent is a distance measure whose suitable choice is discussed. It is shown how sequentially obtained cluster sets can be improved by reclustering, and how items considered as outliers can be removed. The method’s feasible applicability to text analysis is shown.