Study of the impact of different categorical feature encoding techniques on cluster structures

Categorical features are a common type of data used in data analysis, but their non-metric nature makes it difficult to apply standard clustering algorithms. The relevance of the study is conditioned by the need to assess the impact of different methods of recoding (digitisation) of such features on the effectiveness of cluster analysis. The purpose of the study was to investigate how different techniques of categorical data processing affect the quality and structure of clusters. The methodology included the implementation of three models with different approaches to variable coding: without taking into account domain specifics, considering the content of the features, and with alternating the order of application of clustering and dimensionality reduction approaches. LabelEncoder, OrdinalEncoder, One-Hot Encoding, Mapping, and MultiLabelBinarizer were used for coding. In each of the models, clustering was performed using two algorithms – K-Means and agglomerative clustering, which allowed comparison of their sensitivity to changes in data representation. The t-SNE dimensionality reduction method was used to visualise the cluster structure in two-dimensional space. The quality of clustering was evaluated using the Silhouette Score, Dunn Index, Davies-Bouldin Index, and CalinskiHarabasz Index metrics. The data for the analysis were obtained from an open source and contained information about the psycho-emotional state of students. The study found that the basic recoding of categorical features without considering their semantics and context negatively affected the quality of clustering, reducing the accuracy of the division and complicating the interpretation of the results. Instead, the use of domain-oriented coding approaches ensured the development of clusters with clearer boundaries and a more logical internal structure. In addition, it was found that changing the sequence of clustering and dimensionality reduction affects the preservation of local relationships in the data. It was analysed that different approaches change both the number and quality of clusters, which was reflected in the values of the evaluation metrics. The practical significance of the results lies in the possibility of their application by data analysts and machine learning specialists to improve the accuracy of segmentation of complex categorical data