Position regularized Support Vector Domain Description

Support Vector Domain Description (SVDD) is an effective method for describing a set of objects. As a basic tool, several application-oriented extensions have been developed, such as support vector clustering (SVC), SVDD-based k-Means (SVDDk-Means) and support vector based algorithm for clustering data streams (SVStream). Despite its significant success, one inherent drawback is that the description is very sensitive to the selection of the trade-off parameter, which is hard to estimate in practice and affects the extensive approaches significantly. To tackle this problem, we propose a novel Position regularized Support Vector Domain Description (PSVDD). In the proposed PSVDD, the complexity of the sphere surface is adaptively regularized by assigning a position-based weighting to each data point, which is computed according to the distance between the corresponding feature space image and the mean of feature space images. To demonstrate the effectiveness of the proposed PSVDD, we apply the position-based weighting to improve two important clustering extensions, i.e., SVC and SVDDk-Means, which respectively result in two new clustering approaches termed PSVC and PSVDDk-Means. Experimental results on several real-world data sets validate the significant improvement achieved by PSVC and PSVDDk-Means.

► We point out that SVDD is sensitive to the selection of the trade-off parameter.
► We propose a position-based weighting to eliminate the trade-off parameter.
► A new domain description method termed PSVDD is proposed.
► Two new clustering extensions of PSVDD are proposed.