Learning the ‘true’ number of clusters in a given data set is a fundamental and largely unsolved problem in data analysis, which seriously affects the identification of customer segments in marketing research.
In this paper, we discuss the properties of relevant criteria commonly used to estimate the number of clusters. Moreover, we outline two adaptive clustering algorithms, a growing
-means algorithm and a growing self-organizing neural network. In the empirical part of the paper, we find that the first algorithm stops growing with exactly the number of clusters that we get when determining the optimal number of clusters by means of the
-criterion. This cluster solution proves to be rather similar to the one we obtain by applying the neural network approach. To evaluate the clusters, we use association rules. By testing these rules, we show the differences of patterns underlying particular market segments.