Boundary sample extraction support vector data description: a novel anomaly detection method for large-scale data

Abstract

Listen

Translate article

Abstract Support vector data description (SVDD) has been effectively used in many anomaly detection problems. As equipped with kernel functions, its training complexity grows exponentially with the increase of training data, which makes it less practical for large-scale dataset. In this paper, we proposed a boundary samples extraction SVDD(BSE-SVDD) anomaly detection method based on data reduction, aiming for the large-scale data. Firstly, the BSE mechanism is established on the basis of demonstrating that the spatial position of the sample is related to its corresponding Lagrange multiplier. Then, the BSE mechanism is used to search for the local optimal solution of the Lagrange multiplier, and all the training samples are sorted. Finally, the top p of ranked samples are extracted as boundary samples for training, while most of the training samples that may be Non-SVs (NSVs) are removed. Compared with SVDD studies based on data reduction, the experimental results on large-scale datasets show that BSE-SVDD can obtain comparable classification accuracy with greatly improved training speed.