One-Class Support Vector Machines Revisited

The task of One-Class Classification (OCC) is to characterise a single class that is well described by the training data and distinguish it from all others; this is in contrast to the more common approach of binary classification or multi-class classification, in which all classes are well described by the training data. One-class support vector machine algorithms such as OCSVM and SVDD have been shown to be successful in many applications. From our review of the literature, it has emerged that the Gaussian kernel consistently works well in practical applications. Other researchers have shown that OSCVM and SVDD are equivalent under the transformation implied by the Gaussian kernel. A major source of confusion for OCSVM is in how it separates the target data from the origin where the outliers are supposed to lie. In this paper, we review the OCSVM algorithm and we alleviate this source of confusion by proposing a geometric motivation for the OCSVM principle based on separating the target data from the rest of the space, when a Gaussian kernel is used.