Mixture of hyperspheres for novelty detection

Abstract

In this paper, we present a mixture of support vector data descriptions (mSVDD) for one-class classification or novelty detection. A mixture of optimal hyperspheres is automatically discovered to characterize data. The model includes two parts: log likelihood to control the fit of data to model (i.e., empirical risk) and regularization quantizer to control the generalization ability of model (i.e., general risk). Expectation maximization (EM) principle is employed to train our proposed mSVDD. We demonstrate the advantage of the proposed model: if learning mSVDD in the input space, it simulates learning a single hypersphere in the feature space and the accuracy is thus comparable, but the training time is significantly shorter.