DriftWeed A visual metaphor for interactive analysis of multivariate data

We present a visualization technique that allows a user to identify and detect patterns and structures within a multivariate data set. Our research builds on previous efforts to represent multivariate data in a two-dimensional information display through the use of icon plots. Although the icon plot work done by Pickett and Grinstein is similar to our approach, we improve on their efforts in several ways. Our technique allows analysis of a time series without using animation; promotes visual differentiation of information clusters based on measures of variance; and facilitates exploration through direct manipulation of geometry based on scales of variance. Our goal is to provide a visualization that implicitly conveys the degree to which an element's ordered collection (pattern) of attributes varies from the prevailing pattern of attributes for other elements in the collection. We apply this technique to multivariate abstract data and use it to locate exceptional elements in a data set and divisions among clusters.