Visualizing high dimensional feature space for feature-based information classification

Abstract

Feature-based approaches represent an important paradigm in content-based information retrieval and classification. We present a visual approach to information retrieval and classification by interactively exploring the high dimensional feature space through visualization of 3D projections. We show how grand tour could be used for 3D visual exploration of high dimensional feature spaces. Points that represent high dimensional feature observations are linearly projected into a 3D viewable subspace. Volume rendering using splatting is used to visualize data sets with large number of records. It takes as input only aggregations of data records that can be calculated on the fly by database queries. The approach scales well to high dimensionality and large number of data records. Experiments on real world feature datasets show the usefulness of this approach to display feature distributions and to identify interesting patterns for further exploration.