Using Entropy-Related Measures in Categorical Data Visualization

A wide variety of real-world applications generate massive high dimensional categorical datasets. These datasets contain categorical variables whose values comprise a set of discrete categories. Visually exploring these datasets for insights is of great interest and importance. However, their discrete nature often confounds the direct application of existing multidimensional visualization techniques. We use measures of entropy, mutual information, and joint entropy as a means of harnessing this discreteness to generate more effective visualizations. We conduct user studies to assess the benefits in visual knowledge discovery.

1 Introduction

Categorical datasets include survey results in health/social studies, bank transactions, online shopping records, and taxonomy classifications. Such data usually contain a series of categorical variables (i.e., dimensions) whose values comprise a set of discrete categories, such as transaction types, county/town names, product codes, species characters, etc. High dimensional categorical data impose significant challenges for information visualization due to their unique data discreteness. Although major advances have been made on high dimensional data visualization, many successful visualization methods are often undermined when directly applied to categorical datasets.