半监督学习中的几种协同训练算法-南京大学.pdfVIP

半监督学习中的几种协同训练算法 高原 (南京大学 计算机科学与技术系, 南京 210093) A Survey of Semi-Supervised Algorithms with Co-Training Paradigm * Yuan Gao (Department of Computer Science and Technology, Nanjing University, Nanjing 210093, China) Abstract : One can split the Machine Learning into three important research area s, they are Supervised, Unsupervised and Semi -Supervised Learning. As a mixture of Supervised and Unsupervised method, SSL can train with a few of labeled and amount of unlabeled instances. Here, we focus our discussion on Semi -Supervised classification. Since it is proposed in 1998 by BlumMitchell, Co-Training has grown into one of the most important train styles[1]. It builds two or more classifier s which work together during the whole process of training. This paper discusses various important approaches to semi -supervised learning with Co-Training paradigm such as Co-Training, Tri-Training, CoForest and CoTrade. Key words: Semi-Supervised; Co-Training; classifier; labeled; unlabeled; Algorithms 摘 要: 我们可以把机器学习划分为三个重要的研究领域：监督，无监督和半监督学习。作为监督和无监督学 习的混合体，半监督学习能够利用少量标记数据和大量未标记数据进行训练。这里我们主要讨论的是的半监督学 习的分类问题。自1998 年Blum 和Mitchell 提出协同训练算法以来，协同训练风范已经成长为半监督学习中最重 [1] 要的风范之一 。在整个训练过程中它建立起两个或两个以上的分类器并让他们协同工作。本文主要讨论了一些 在半监督学习中具有协同训练风范的算法：Co-Training，Tri-Training，CoForest 和CoTrade 。 关键词: 半监督;协同训练;分类;标记;未标记;算法 中图法分类号: TP301 文献标识码: A 1 引言 分类问题的核心是如何训练出一个分类器，它的输入是一个待分类示例的特征向量，对应的输出是一个 ∑ ( ) ( ) ( ) | ( ) ( ) 合理的类别标记。贝叶斯公式 P | = 表明我们可以将先验概率P 转换为后验概率P | 。 ( )