基于三支决策的广义代价敏感近似属性约简  

Generalized cost-sensitive approximate attribute reduction based on three-way decisions

作　　者：方宇[1] 高磊[1] 刘忠慧[1] 杨新 Fang Yu;Gao Lei;Liu Zhonghui;Yang Xin(School of Computer Science,Southwest Petroleum University,Chengdu 610500,China;School of Economic Information Engineering,Southwestern University of Finance andEconomics,Chengdu 611130,China)

机构地区：[1]西南石油大学计算机科学学院,四川成都610500 [2]西南财经大学经济信息工程学院,四川成都611130

出　　处：《南京理工大学学报》2019年第4期481-488,共8页Journal of Nanjing University of Science and Technology

基　　金：国家自然科学基金(41604114);西南石油大学2018年高等教育教学改革研究项目(X2018JGYB043\37\38)

摘　　要：在粗糙集领域,属性约简的首要任务是在保持原有数据相关特性的前提下,获取一个最小的属性子集。代价敏感学习的目标旨在最小化各类代价。而近似属性约简的意义在于让决策者能够权衡代价承受能力和知识发现的程度。本文在定性和定量的标准下提出了代价敏感近似属性约简的问题;定性标准指不可分辨能力,定量标准指近似参数ε和代价。基于三支决策和可分辨矩阵,提出了解决代价敏感近似属性约简问题的框架:首先,定义了属性子集的质量函数,该函数解释了多粒度结构;其次,通过考察属性重要度,提出了“性价比”指标的适应函数;进而利用提出的适应函数和三支决策中的(α,β)阈值对三分属性集合;最后,设计了删除策略和添加策略的代价敏感属性约简算法。从实验结果分析上验证了算法的有效性,体现了提出的问题和理论框架具有更广义的解释和适应性。In the research spectrum of rough sets,the primary task of attribute reduction is to obtain a minimum subset of attribute set while maintaining the relevant features of the original data.The goal of cost-sensitive learning is to minimize the various costs.The significance of approximate attribute reduction is to enable decision makers to leverage the cost tolerance and the grade of knowledge discovery.This paper proposes a cost-sensitive approximate attribute reduction problem with both qualitative and quantitative criteria.The qualitative criteria refers to(in)discernibility,and the quantitative criteria refers to approximate parametersεand costs.Based on the three-way decisions and discernible matrix,this paper portrays a framework to solve the problem of cost-sensitive approximate attribute reduction.First,we define the quality function of attribute set which explains the multi-granularity structure.Second,we propose a‘cost-performance index’fitness function to evaluate the importance of attribute,then the proposed fitness function and(α,β)thresholds pair with three-way decisions are applied to tri-partition the attribute sets.Finally,we design the two algorithms(deletion-based and addition-based)to tackle the reduction problem.The validity of the algorithms is verified by the experimental result analysis,which proves that our framework has broader adaptability and applicability.

关 键 词：属性约简 代价敏感学习 (不)可分辨能力 粒计算 三支决策 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]