Ordering Categorical Data to Improve

| Visualization provides a means for exploratory analysis of large scale, complex data. In domains such as network management, these data often have categorical attributes, such as host names and event types. Unfortunately, large scale visualiza-tions of categorical data are diicult to construct since categorical values have no inherent order. We consider two visual tasks: nding groups of similar values of a single categorical attribute and nding relationships between values of diierent categorical attributes. To assist in these tasks, we develop an algorithm that orders categorical values using the following steps: (1) construct natural clusters of categorical values based on domain semantics (e.g., group together hosts that emit events at the same time); (2) order the clusters; and (3) order the categorical values within each cluster. We demonstrate the beneets of this approach by applying it to scatter plots and parallel coordinate plots of event data collected from a corporate Intranet. I. Introduction Visualization has become increasingly important for the analysis and exploration of multidimensional data (e.g., 5]]1] 3]]6]). While a great deal of work has addressed visualization of numeric data, many domains require visualization of large amounts of categorical data. Examples of such data are: city names in census data, equipment manufacturers in inventory data, and host names in network management data. Unlike numeric data, categorical values do not have an order. This is problematic for commonly-used visu-alization techniques, such as scatter plots and parallel coordinate plots, since categorical values need to be mapped to axis coordinates. Technically, any order of the categorical values is valid. However, as we will show later using data collected from a corporate In-tranet, properly ordering the data can greatly improve the quality of the visualizations. Although ordering (or permuting) categorical values is known to be important for visual analysis of categorical dataa6], we are aware of only two approaches that have been applied to plots. The rst orders values manually. This requires a knowledgable end-user, and