Visual Clustering and Classiication: the Oron- Say Particle Size Data Set Revisited 2.2 Parallel Coordinate Plots 2.3 Linked Low{dimensional Views 3 the Oronsay Particle Size Data Set


Interactive statistical graphics can be eeectively used to nd natural groupings in observations. In this paper we want to demonstrate how clustering and classiication can be done with three approaches based on highly interactive graphical environments: high{dimensional scatterplots as available in XGobi, parallel coordinate plots as available in ExplorN, and linked low{ dimensional views as available in Manet. We will point out the strengths and the weaknesses of these techniques by comparing their behavior when applied to the Oronsay particle size data set. 2 1 Motivation Clustering and classiication have always been highly related to the geometric structure of a data set. Visualization of that structure has been a main goal. Graphical devices have been widely used to support this visualization, but typically the use of graphics is restricted to the presentation of results and not to the exploration of clusters. In the last three decades many display types have been proposed for analyzing multivariate data and almost all creators of a new plot type claim that their method is powerful in clustering and classiication. So, why do these graphical techniques play a minor role in the classiication community? A main reason might be that classiication using these plots is hard to achieve in a static environment. The rise of high interaction graphics provides the user with the exibility and power needed to detect and assign clusters. The outline of the paper is as follows. In Section 2 we describe the basic notions of the high interactive techniques that we use and relate them to a dynamic graphics program: scatterplot rotation with projection pursuit and grand tour (XGobi), parallel coordinate plots combined with a d{dimensional grand tour (ExplorN), and linked low{dimensional views with user speciied subsetting (Manet). The Oronsay particle size data set, used throughout this paper, will be introduced in Section 3. In Sections 4, 5, and 6, we describe our analyses in XGobi, ExplorN, and Manet, respectively. We conclude with a summary in Section 7. Analyses by high{interaction color graphics do not translate well into static grayscale pictures in a paper{based publication. For this reason, all of the images referred to in the following sections plus additional material to clarify intermediate steps are available in full color on our webpage: High interaction graphics have been developed within the last 15 years to enhance the graphic facilities developed for exploratory data analysis since the early 1960's. Eick & Wills …

2 Figures and Tables

Cite this paper

@inproceedings{Wilhelm1999VisualCA, title={Visual Clustering and Classiication: the Oron- Say Particle Size Data Set Revisited 2.2 Parallel Coordinate Plots 2.3 Linked Low\{dimensional Views 3 the Oronsay Particle Size Data Set}, author={Adalbert F. X. Wilhelm and Edward J. Wegman}, year={1999} }