Abstract: A method for extracting data which most probably are suport vectors for SVM by the Mahalanobis distance from a vector to a class is presented. How to compute Mahalanobis distance in the input and feature space is described in detail. The algorithm is fast since there are efficient methods for finding eigenvalues and eigenvectors of a symmetric matrix or comptuing pseudoinversion involved in finding the Mahalanobis distance. The training time for SVM can be reduced when the training set is preprocessed in this way. Experimental results illustrate its effectiveness.

Key words: support vector machine, support vector, Mahalanobis distance, kernel function, pseudoinvertion