Comparison of different methods of categorical data cluster analysis

This diploma thesis deals with methods of categorical data cluster analysis. Many categorical data clustering methods have not been properly researched because these methods are still developing. The aim of this work is to present, explore and compare three selected methods of categorical data clustering, one of them is hierarchical clustering using the Gower coefficient as a representative of classical methods, and the other two, ROCK and COOLCAT algorithms, as representatives of new approaches. The analytical part of the thesis deals with the comparison of methods on seven real data sets, which contain one main class variable. The results show that the most accurate classification of the categories of the class variable into the created clusters and the highest quality clusters of most data sets has clustering using the ROCK algorithm after removing unclassified observations, which the algorithm considered as outliers. Nevertheless, this method is not always optimal, because it often loses information about part of the observations. If it is required to classify all observations, the best clusters were created for most data sets using the COOLCAT algorithm.