Efficient Dimensionality Reduction on Undersampled Problems through Incremental Discriminative Common Vectors

An efficient incremental approach to the discriminative common vector (DCV) method for dimensionality reduction and classification is presented. Starting from the original batch method, an incremental formulation is given. The main idea is to minimize both matrix operations and space constraints. To this end, an straightforward per sample correction is obtained enabling the possibility of setting up an efficient online algorithm. The performance results and the same good properties than the original method are preserved but with a very significant decrease in computational burden when used in dynamic contexts. Extensive experimentation assessing the properties of the proposed algorithms with regard to previously proposed ones using several publicly available high dimensional databases has been carried out.