误差在Cost-Sensitive分类中的应用  

Application of Errors in Cost-Sensitive Classifications

作　　者：廖元秀[1] 周生明[1] 

机构地区：[1]广西师范大学计算机科学与信息工程学院,广西桂林541004

出　　处：《广西师范大学学报（自然科学版）》2011年第2期110-113,共4页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金资助项目(60963008)

摘　　要：针对使检查代价和误分类代价最小化的Cost-Sensitive学习,讨论误差在分类过程中的应用,提出一个带阈值的决策树,并给出一个带阈值的检查策略。在基于Cost-Sensitive学习的分类中,确定属性值所用到的检测手段和设备精度存在一定的误差值,评估误分类代价更是有较大的误差。另外,很多分类问题并不要求达到百分之百的正确率,允许有一定的误差范围。把这些误差的边界看作是一个阈值,利用这种阈值来简化决策树的建立,改进检查策略的设计,提高分类效率。Aiming at the minimization problem of test costs and misclassification costs on Cost-Sensitive learning,application of errors in classifications are discussed. A kind of decision trees and test strategies with thresholds are proposed. There are errors on both Cost-Sensitive classifications resalted from meth- ods of tests and equipment accuracy and evaluating misclassification. In addition, many classification problems are not required to achieve one hundred percent classification accuracy. The boundaries of these errors are regarded as a kind of threshold values. The establishment of decision trees is simplified and the design of test strategies and the classification efficiency are improved by using these threshold values.

关 键 词：Cost-Sensitive学习 分类代价 检查策略 分类误差范围 阈值 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]