Preserving Class Discriminatory Information by Context-sensitive Intra-class Clustering Algorithm

Many powerful techniques in supervised learning (e.g. linear discriminant analysis, LDA, and quadratic classifier) assume that data in each class have a single Gaussian distribution. In reality, data in the class of interest, i.e., the object class, could have non-Gaussian distributions and could be isolated into several subgroups by the data from other classes (the context classes). To address this problem, one possible way is to partition one class into several subclasses. This intra-class clustering should depend on the data structure of the class of interest (object class) as well as the distributions of all other classes (context classes). In this paper, we presented a novel method of intra-class clustering which can divide a nonGaussian class data into several Gaussian-like clusters, and at the same time this algorithm is context sensitive, which can maximally reduce the overlapping among resulting classes and also between the object class and the context classes. The method can serve as a general 1 data preprocessing method to improve performance of supervised learning algorithms such as LDA and quadratic classifiers.