Active learning based support vector data description method for robust novelty detection

摘要

Practical industrial data usually has non-Gaussian data distribution and nonlinear variable correlation. Because support vector data description (SVDD) has no Gaussian limitations and can be extended to the nonlinear case by applying the kernel trick, it is one of the most widely used novelty detection methods. However, there is a great deal of actual industrial data that are mixed with much noise and uncertainty. Furthermore, SVDD may perform worse when the amount of data is too large and the data quality is poor. Describing the whole distribution with a small number of labeled samples has great practical significance and research value. This paper proposed an active learning-based SVDD method for robust novelty detection. It can reduce the amount of labeled data using an active learning framework, generalize the distribution of data and reduce the impact of noise by using the local density to guide the selection process. Experiments on two-dimensional synthetic distributions, UCI datasets and the Tennessee Eastman Process (TEP) show the effectiveness of the proposed method.