Measure-Driven Visual Analytics of Categorical Data

Abstract

Visual Analytics (VA) enables data analysts and domain experts to engage in analytical reasoning through interactive visual interfaces. One type of data often encountered in data analysis tasks is categorical data. Unlike numerical data, categorical data with nominal attributes has no inherent order or scale and, therefore, does not lend itself to the application of common arithmetic operations. However, many data mining and visualization techniques are predominantly based on numerical data. Notwithstanding these challenges, the analysis of categorical data is crucial in various domains, including linguistics and software engineering. This dissertation addresses the challenges posed by categorical data, including difficulties in establishing an order of attributes for visualization and defining numerical abstractions. This work bridges the qualitative-quantitative divide in the visual analysis of categorical data by introducing abstractions that improve the readability of categorical data visualizations, developing new strategies for applying methods typically designed for numerical data, and exploring their interplay with numerical data. This thesis is structured in three parts: The first part introduces quality measures for the Parallel Sets visualization. In addition, we present measures that guide the exploration of categorical data projections by suggesting attributes that differentiate groups of data items. The second part presents measure-driven approaches for expressing categorical data properties and deriving numerical representations for the domains of linguistics and software engineering, demonstrating the power of measure-driven approaches in real-world applications. The third part addresses the joint analysis of categorical attributes and numerical data dimensions. It offers strategies for the use of categorical data for model training and exploratory data analysis in supervised and unsupervised frameworks. Finally, this thesis outlines the limitations and lessons learned from the explored measure-driven approaches and suggests future directions for more effectively integrating categorical data into VA with the goal of improving the readability of visualization, pattern quantification and user guidance. In conclusion, this work improves the analysis and visualization of categorical data by proposing new measure-driven approaches, improving readability and interpretability of visualizations, providing domain-agnostic and domain-specific support for exploratory data analysis, and their integration into supervised and unsupervised VA frameworks.