中文摘要：

以统计学习理论为背景,以核方法为基础的两类典型单类分类算法：单类支持向量机（OCSVM）和支持向量数据域描述（SVDD）,均以降低VC维为目标,其中前者通过寻找一个远离原点的超平面,使目标数据所在的正半空间尽量最小;而后者通过寻找一个包含大部分目标数据的最小超球,实现体积最小化.围绕上述两算法,已有大量改进形式出现.本文以此为主线,分别从模型构建、模型改进和数据预处理的角度,进行了回顾和阐述,并对各算法的特点给出了相应的总结.

英文摘要：

As state-of-the-art algorithms based on kernel method, one-class SVM （OCSVM）and Support Vector Data De- scription （SVDD） root into the sound theoretical basis of statistical learning theory. In order to decrease the VC dimen- sion for promoting the generalization ability, OCSVM tries to find a hyperplane with the furthest distance to the origin for minimizing the positive half space lived by most of the target data; While SVDD tries to find the minimal volume hyper- sphere enclosing most of the given samples. Focusing on the two algorithms, some variants or improved versions are pro- posed to avoid some disadvantages of the above models. In this paper, we review most of these variants and give a detailed relation among the discussed algorithms to the original models.