One-Class分类器研究

摘要

In Traditional binary or multiple classification problem of machine learning, each class of samples is necessary for classifier design;however,in some case only one class of samples can be acquired (due to the complexity or the expensive costs), so we have to learn from the only one class of samples and form the data description for classification. The classification problem is named as one-class classification. Since now there exist the domain-specific and the generic methods, this paper first presents an overview on one-class classification, then emphasizes on the analysis of the kernel-based one-class classifiers and divides this class of methods into two types, that is, dual-based and kernel-induced distance based. Hereafter, the characteristics of these two types of methods are analyzed. Finally, we summarize the implementation techniques and applications of one-class classification in fault analysis,anomaly detection,and disease diagnosis and hostility recognition.%传统的多类分类问题需要多类样本训练分类器,然而由于样本获取(复杂性或代价)的原因很多情况下只能获取一类样本,故只能利用这一类样本进行学习,形成数据描述从而实现分类,故称之为One-Class分类.鉴于目前单类分类研究存在领域相关方法和通用方法百花齐放的格局,本文首先给出了当前One-Class分类器的研究综述,然后重点针对基于核方法的单类分类器进行分析,将该类方法分为对偶方式和核距离方式两类,并分析各自的特点.最后论文介绍了目前的单类分类器的应用领域,指出其在故障分析、异常检测、疾病诊断和敌我识别等现实问题中的重要作用.