Online anomaly detection with an incremental centred kernel hypersphere

Anomaly detection is an important aspect of data analysis. Kernel methods have been shown to exhibit good anomaly detection performance, however, they have high computational complexity. When anomaly detection is performed on a data stream, computational complexity is a key issue. Our approach uses the kernel hypersphere, which does not require a computationally complex operation in order to form the model. We introduce an incremental update and downdate to the model to further reduce computational complexity. Evaluations on synthetic and real-world datasets show that the incremental kernel hypersphere exhibits competitive performance when compared to other anomaly detectors.