On the comparison of conceptual clustering and numerical taxonomy

Conceptual clustering requires that clusters formed in the process shall be definable in terms of simple formulas in the predicate calculus. Michalski and Stepp [1] have argued that the results obtained with this method are clearly superior to traditional methods of numerical classification, so that an order of magnitude degradation in performance is acceptable. In this paper the results and comparisons presented by Michalski and Stepp are reviewed and shown to be less than adequate to support such a conclusion. There are considerable problems with data coding and standardization, as well as choice of similarity measure, that make the results difficult to evaluate. Even accepting the clusters, two different valorizing schemes are used to evaluate the results obtained. In addition, the traditional agglomerative algorithm employed in numerical classification procedures can be adapted to perform conceptual clustering without an enormous degradation in performance. However, the value of the method can only be regarded as unproven.