面向新异检测的启发式约减支持向量数据描述

Heuristic reduction support vector data description for novelty detection

针对支持向量数据描述(SVDD)单类分类方法运算复杂度高的缺点,提出一种启发式约减支持向量数据描述(HR-SVDD)方法.以启发的方式从原有训练集中筛选出部分样本构成约减训练集,对约减训练集进行二次规划解算,得到支持向量和决策边界.通过不同宽度系数高斯核SVDD特征的讨论,证明了HR-SVDD的有效性.人工数据集和真实数据集上的实验结果表明,HR-SVDD分类精度与传统支持向量数据描述相当,但具有更快的运算速度和更小的内存占用.

A method of heuristic reduction support vector data description（HR-SVDD） is proposed for speeding up the support vector data description（SVDD） one-class classification method. The HR-SVDD first builds a reduced training set by selecting a portion of samples from training set in heuristic, and then completes the quadratic programming using the reduced training set rather than the original training set. The efficiency of proposed method is demonstrated by discussing characteristic of Gaussian kernel SVDD with different width parameters. For demonstration, experiments on artificial and real-world datasets are conducted, and the results show that the classification accuracy of HR-SVDD is nearly identical to that of conventional SVDD, but with faster running speed and less memory usage.