Abstract: A space support vector domain classifier (SSVDC) is proposed. In the training process, the support vector domain description (SVDD) is applied to both the positive and negative classes, disconnect regions are obtained according to the description boundaries and different classification rules are erected for the corresponding regions. In the test phase, the distances from the test sample to each hypersphere centers are computed, the region that the test samples belong to is confirmed according to the relations between their central distances and the hyperspheres radii, so that corresponding rules can be adopted. Numerical experiments on UCI data show that compared with existing algorithms SVM and SVDC, SSVDC has better robustness, a shorter training time of about 20.6％ SVM and a classification accuracy which is about 45.9％ higher than that of SVDC in the best case.

Key words: space, support vector domain classifier, SVDD, description boundary, region