Issue No.05 - September/October (1999 vol.19)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/38.788795
Most automated clustering algorithms do not work effectively on high-dimensional data -- they are likely to miss clusters with certain unexpected characteristics. For example, the so-called "curse of dimensionality" makes it difficult to find the necessary parameters for tuning the clustering algorithms to the specific application. We propose novel visual mining techniques to overcome these problems. The idea is to support the critical steps of an advanced automated clustering algorithm by visualization techniques. The automated clustering algorithm uses projections of the point density of the high-dimensional data to find good separators between the clusters. The visualization techniques we use allow easy identification of important data characteristics. Since the number of interesting projections may become large, we provide different visualization techniques ranging from abstract iconic representations of the separation potential to pixel-oriented overview plots of the multi-dimensional projections. The visualizations also allow specifying complex hyper-polygonal separators directly within the visualization. This permits finding clusters that no automatic algorithm can determine. We integrated all the visualization techniques using a tree-like visualization of the projection and separator hierarchy. Experiments applied our new visualization techniques in a real application from molecular biology.
Visual Support of the Clustering Process, Visual Data Mining, Pixel-oriented Visualization Techniques, Iconic Visualization Techniques
Daniel A. Keim, Alexander Hinneburg, "HD-Eye: Visual Mining of High-Dimensional Data", IEEE Computer Graphics and Applications, vol.19, no. 5, pp. 22-31, September/October 1999, doi:10.1109/38.788795