Discriminant Subspace Learning Based on Support Vectors Machines

A new method for dimensionality reduction and feature extraction based on Support Vector Machines and minimization of the within-class data dispersion is proposed. An iterative procedure is proposed that successively applies Support Vector Machines on perpendicular subspaces using the deflation transformation in such a way that the within-class variance is minimized. The proposed approach is proved to be a successive SVM using deflation kernels. The normal vectors of the successive hyperplanes contain discriminant information and they can be used as projection vectors for feature extraction and dimensionality reduction of the data. Experiments on various datasets are conducted in order to highlight the superior performance of the proposed algorithm.