MeatballWiki | RecentChanges | Random Page | Indices | Categories
Information visualization is often tasked with visualization very high dimensional data on two or three dimension displays. One technique is to statistically reduce the number of dimensions in the source data.
- PrincipleComponentsAnalysis?. Separates minor "noise" dimensions of high dimensional vectors so they can be ignored, thus reducing the overall dimensionality of the vectorspace to be visualized. Requires a lot of memory and computational power.
- MultiDimensionalScaling? (MDS). Visualizes a high-dimensional vector space by projecting it onto a lower dimensional vector space. It does this by reducing the problem to one of relative similarity between the items, rather than their absolute values, say by comparing vectors by a Euclidean distance.
- SelfOrganizingFeatureMaps?. A neurocomputational model based on competitive, unsupervised learning. First it maps items onto a 2D grid, and then it iteratively improves this mapping until quiessence or halt.
- SimilarityBasedDimensionClustering?. Through a clustering algorith, creates a dimension hierarchy which can be navigated through novel non-Cartesian techniques.
Huang, S., Ward, M. O., Rundensteiner, E. A. (2003). Exploration of dimensionality reduction for text visualization. Technical Report TR-03-14, Worcester Polytechnic Institute, Computer Science Department. Available from http://citeseer.ist.psu.edu/huang03exploration.html