A dimensionality reduction technique for enhancing information context.

Abstract

A computer processing technique is advanced which seeks to retain or improve data information context while reducing the dimensionality of data representation. Defining information context as the relative proximity of data points, a nonlinear transformation is analytically derived which utilizes Euclidean distance to one or more reference points to provide a measure of similarity "between data points. The nonarbitrary reference points are selectively manipulated to provide, given certain constraints, a unique mapping from high dimensional space to one or more dimensions for each point in space. The transformation process enhances class clustering and interclass separation in the lower dimensional representation. Computer processed experimental results are presented of reduction from 32, 10, and 3 space into 2 space for both synthetic and real world data. Utilizing a ratio of intraclass variance to interclass variance as a figure of merit and as one possible optimization criterion, this technique yielded a significant ratio improvement in mapping from higher dimensional space into 2 dimensional space for all cases examined.