Feature Extraction through Rotations

A procedure is developed for obtaining the lower dimensional representation of highdimensional observations stemming from different classes that best distinguishes among the classes. The method finds a low dimensional subspace such that the estimated probability of the projected data belonging to each population is as close to the true assignment as possible. This is achieved by starting from a random subspace and applying successive rotations in the direction of descent of the Kullback-Leibler divergence between the true and computed assignments. The method is applied to classification, comparing its performance with benchmark methods on both synthetic data and on the Wisconsin Breast Cancer Diagnostic dataset.