Alternative visualizations on categorical data in cluster analysis

This master’s thesis deals with proposing a methodology for visualizing cluster analysis of categorical data. While there are many visualization options for quantitative variables, these options are lacking for categorical data. This work provides an overview of possible approaches to visualizing categorical data and cluster analysis. Furthermore, a new visualization methodology is proposed in the work, which addresses issues associated with visualizing categorical data and thus will be a suitable tool for graphically representing the results of cluster analysis. To achieve this, the basic principles of cluster analysis are first described with focus on clustering categorical data. The thesis then discusses various methods that can be used to order categories and determine their distances. Based on the described theoretical knowledge, the thesis proposes new methods named HCADM, HCAKL, LBCADM, and LBCAKL. These methods address the problem of orderig categories on the graph axes by using hierarchical clustering and latent blocks clustering. For the second problem, determining distances between categories, the introduced methods use measures such as OF, IOF, Lin, or the Kullback-Leibler distance. The effectiveness of the HCADM, HCAKL, LBCADM, and LBCAKL methods is examined on two datasets, showing that the graphs generated by these methods can be utilized for visualizing cluster analysis of categorical data. The graphs lead to a better representation of clusters and can thus be used as an equivalent of a cluster scatter plot for categorical variables.