In this paper, we propose a new approach for feature space visualization, preserving a pattern distribution structure. In statistical pattern recognition, it is useful to obtain a two- or three- dimensional projection of the given multivariate data to permit a visual examination of the data. The principal components analysis and the quantification theories are traditional dimensionality reduction techniques for feature space visualization of multivariate data. A two- or three- dimensional mapping using these techniques usually breaks a pattern distribution structure of the data, which is not considered by the projection. On the contrary, we propose a new approach of a dimensional reduction technique preserving a pattern distribution structure analyzed in feature space. And we propose a novel visualization method under our approach, called the cluster discriminant analysis: an optimal linear projection for a cluster structure of the data. In addition, our method is applied to an interface of database retrieval.
展开▼
