New Method of Variable Selection for Binary Data Cluster Analysis

Cluster analysis of binary data is a relatively poorly developed task in comparison with cluster analysis for data measured on stronger scales. For example, at the stage of variable selection one can use many methods arranged for arbitrary measurement scales but the results are usually of poor quality. In practice, the only methods dedicated for variable selection for binary data are the ones proposed by Brusco (2004), Dash et al. (2000) and Talavera (2000). In this paper the efficiency of these methods will be discussed with reference to the marketing type data of Dimitriadou et al. (2002). Moreover, the primary objective is a new proposal of variable selection method based on connecting the filtering of the input set of all variables with grouping of sets of variables similar with respect to similar groupings of objects. The new method is an attempt to link good features of two entirely different approaches to variable selection in cluster analysis, i.e. filtering methods and wrapper methods. The new method of variable selection returns best results when the classical k-means method of objects grouping is slightly modified.