代价敏感代表选举的邻域覆盖粗糙集分类方法  

Cost-sensitive representatives-based classification through covering-based neighborhood rough set

作　　者：刘福伦[1] 闵帆[1] 张本文[2] 

机构地区：[1]西南石油大学计算机科学学院,成都610500 [2]四川民族学院计算机科学系,康定626001

出　　处：《江苏科技大学学报（自然科学版）》2017年第2期190-195,共6页Journal of Jiangsu University of Science and Technology:Natural Science Edition

基　　金：国家自然科学基金资助项目(61379089);四川省教育厅自然科学基金资助项目(13ZA0136)

摘　　要：基于邻域的覆盖粗糙集已经被用于构建数值型数据分类器.文中将该理论应用于针对名词性数据的代价敏感分类器.首先,用对象间的相似关系确定每个对象的相似度阈值,并通过该阈值确定每个对象的邻居以及邻域;然后,通过覆盖约简算法选举出一组称为代表的对象;最后,当需要对新对象进行分类时,计算它与各代表之间的曼哈顿距离,并综合误分类代价向量制定不同分类策略,以最小化误分类代价期望值.在UCI数据库中的数据集之上与主流的代价敏感ID3和C4.5算法对比,结果表明新算法能取得更好的分类效果.Covering-based neighborhood rough set has been employed to design classifiers especially for numerical data. In this paper,we apply the theory to build a cost-sensitive classifier for nominal data. First,we determine the similarity threshold for each instance according to the similarity relation with others; with this similarity threshold,neighbors are determined and constructed. Second,a covering reduction algorithm is designed to find out a set of instances called representatives as the basis of the classifier. Third,while classifying a new instance,the Manhattan distances and misclassification cost vector are applied to minimize the expected misclassification cost. Results on some UCI datasets show that our new algorithm is generally comparable to or better than costsensitive ID3 and cost-sensitive C4. 5.

关 键 词：代价敏感 代表 邻域 相似度 距离 覆盖 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]