Multi-sphere Support Vector Data Description for

SVDD has been proved a powerful tool for outlier de-tection. However, in detecting outliers on multi-distribution data, namely there are distinctive distributions in the data, it is very challenging for SVDD to generate a hyper-sphere for distinguishing outliers from normal data. Even if such a hyper-sphere can be identified, its performance is usu-ally not good enough. This paper proposes an multi-sphere SVDD approach, named MS-SVDD, for outlier detection on multi-distribution data. First, an adaptive sphere detec-tion method is proposed to detect data distributions in the dataset. The data is partitioned in terms of the identified data distributions, and the corresponding SVDD classifiers are constructed separately. Substantial experiments on both artificial and real-world datasets have demonstrated that the proposed approach outperforms original SVDD.1