Abstract: To accelerate the training of SVDD, a support vector domain description method RSVDD based on reduced sets is proposed. Since the boundary is determined by a small portion of data called support vectors which distribute around the description boundary; the proposed algorithm treats the distance to the center as a probability measure of support vectors for each sample, and selects the former ones ranking as the reduced sets to participate in the SVDD training. Simulations over artificial and benchmark data show its effectiveness and superiority: RSVDD reduces the training scale and the training time, while maintaining the accuracy of targets and outliers.

Key words: support vector domain description, reduced sets, central distance, support vectors