基于粗糙集理论的协同训练算法  

Novel co-training algorithm based on rough sets

作　　者：盛小春[1] 岳晓冬[2,3] 

机构地区：[1]江苏理工学院云计算与智能信息处理常州市重点实验室,江苏常州213001 [2]上海大学计算机工程与科学学院,上海200444 [3]同济大学计算机科学与技术系,嵌入式系统与服务计算教育部重点实验室,上海201804

出　　处：《计算机应用研究》2013年第12期3546-3550,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61103067);常州市云计算与智能信息处理重点实验室资助项目(CM20123004);江苏理工学院青年基金资助项目(KYY11093)

摘　　要：为了提升风险决策环境下协同训练的效果,提出了一种基于粗糙子空间的协同决策算法。首先利用粗糙集属性约简的概念,将部分标记数据属性空间分解为两差异性较大的粗糙子空间;在各子空间上训练分类器,并依据各分类器决策风险代价及隶属度将无标记数据划分为可信、噪声和待定样本。综合两分类器的分类结果,标注少量可信无标记样本后重复协同训练。从理论上分析了算法性能提升的区间界,并在UCI数据集上进行实验,验证了模型的有效性及效率。In order to improve the performance of co-training in the context of decision with risk, this paper proposed a rough subspace-based co-training algorithm. Based on the concept of attribute reduction in rough sets, the algorithm first splitted all condition attributes of partially labeled data into two diverse rough subspaces. Then the two classifiers trained from derived rough subspaces and classified the unlabeled data into confident, noise and uncertain samples with consideration of the classification risk and membership of decision class. Finally,it labeled a few of confident samples for the two classifiers to learn from each other in iterative manner. It theoretically analyzed the performance of proposed algorithm, and empirical results on selected UCI data sets also show its effectiveness.

关 键 词：协同训练 属性约简 粗糙集 粗糙子空间 决策风险 

分 类 号：TP182[自动化与计算机技术—控制理论与控制工程]