Dynamic hypersphere SVDD without describing boundary for one-class classification

Abstract

Support vector data description (SVDD), an efficient one-class classification method, captures the spherically shaped boundary around the same class data and achieves classification for setting the boundary related to support vectors (SVs). As SVDD constructs an irregular hypersphere in high-dimensional space, it is unreasonable to keep the classification boundary a constant value. When the classification dataset is complicated, constant classification boundary will decrease the accuracy of classification. In this paper, we present a dynamic hypersphere SVDD (DH-SVDD) without describing boundary for one-class classification. In training process, important SVs of training dataset describe the static hypersphere. In testing process, dynamic hypersphere is described according to the new important SVs of the testing sample and training dataset. If there is a significant change of hypersphere structure, it means the new sample is an outlier. In this method, without any classification boundary, it can complete one-class classification with fully considering the related information of new sample and historical dataset. Thus, it can significantly improve the one-class classification accuracy of SVDD in complex datasets. Comparison is conducted among the proposed DH-SVDD, K-chart SVDD, Max limit SVDD and Validation limit SVDD. The effectiveness of the proposed method is also verified by the experimental UCI datasets.