Difficulties of Cluster Analysis on Mixed-Type Variables

Working with real data often involves clustering of mixed-type variable datasets, which is followed by certain difficulties resulting from lack of guidelines on the usage of clustering methods for mixed-type data. Instead of choosing the optimal method for mixed data clustering, this study aims to determine on which conditions do the specialized cluster analysis methods - hierarchical clustering using Gower's distance and ${k}$-prototypes algorithm, produce similar partitioning results. Resulting clusterings were after compared with clusters attained by the usage of one-hot encoding technique for transforming categorical data into numerical values. To determine the degree of similarity between methods special metrics were used. Data used in this study is real data obtained from real estate sales advertisements in Saint Petersburg.