FAST TRAINING OF SVDD BY EXTRACTING BOUNDARY TARGETS
Junyu Liang, Shuo Liu, Dengfeng Wu · Iranian journal of electrical and computer engineering · 2009
Training support vector domain description (SVDD) involves solving a constrained convex quadratic programming, which requires large memory and enormous amounts of training time for large-scale data, an extraction strategy is proposed to extract boundary targets, based on the observation that the description boundary is determined by a small subset of training data called support vectors. Namely, the number of samples that scatter around each training target is calculated and taken as the measure of nearness to boundary targets, according to which the training samples are ranked in ascending order. Those former ranked samples are extracted as the boundary targets and are used for SVDD training. We compare the effectiveness of the proposed SVDD using extraction strategy with SVDD in terms of training accuracy, training scale and training time on artificial and benchmark data. Numerical experiments show that: the training scales and the training times are reduced without any loss of accuracy, which therefore can be used for large- scale data.