Iterative character weighing in numerical taxonomy.

Abstract Classical numerical taxonomic methods are extended as to include character weighing. Character weights are assigned iteratively by virtue of the extent to which the character supports the clustering found in the previous cluster analysis (based on different character weights). The behaviour of agglomerative clustering methods, while character weights are thus adjusted, is investigated in this paper. This behaviour is quite surprising: in the course of the iteration the cluster structure becomes more pronounced and new, clearly distinguishable, clusters emerge. These clusters prove to be well interpretable in terms of classical taxonomic groups. This result suggests that concept formation as done in classical taxonomy (and probably elsewhere as well) may be modelled as a feedback of more global information, which is obtained from local similarities, onto the evaluation of local similarities.