局部视角下的邻域粒度及其属性约简  

Neighborhood granularity and attribute reduction from a local perspective

作　　者：徐天喜 宋晶晶[1] 陈建军 徐泰华 XU Tianxi;SONG Jingjing;CHEN Jianjun;XU Taihua(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212003,China;Key Laboratory of Marine Big Data Mining&Application,Zhejiang Ocean University,Zhoushan 316022,China)

机构地区：[1]江苏科技大学计算机学院,镇江212003 [2]浙江海洋大学海洋大数据挖掘与应用重点实验室,舟山316022

出　　处：《江苏科技大学学报（自然科学版）》2024年第5期92-99,共8页Journal of Jiangsu University of Science and Technology:Natural Science Edition

基　　金：国家自然科学基金项目(62006099,62076111);浙江省海洋大数据挖掘与应用重点实验室开放课题(OBDMA202104)。

摘　　要：在基于粗糙集的粒计算中,邻域粒度是常见的信息粒化表现之一.邻域粒度在不同条件属性之间存在一定差异,如果条件属性生成的邻域粒度越细,那么该条件属性对样本的区分性越强.条件属性在全局样本上生成的邻域粒度,即在全局视角下,会存在一些局限性,例如多个条件属性可能会生成相同或相似的邻域粒度,这样不利于对条件属性作进一步区分.为了弥补这样的缺陷,首先从局部视角出发,提出了局部邻域粒度的概念.与全局视角生成的邻域粒度不同,计算局部邻域粒度首先对全部样本进行分割,然后对分割后的样本分别计算邻域粒度.其次,基于局部邻域粒度提出了一种属性约简改进算法,该算法主要思想是以局部邻域粒度为准则,在选择条件属性时剔除对样本区分性较弱的条件属性.在UCI数据集上选取12组数据集,在两种不同的度量条件下,提出的算法与另外两种算法进行对比实验.实验结果表明,与另外两种算法相比,文中算法在分类准确率和分类稳定性上都有明显优势.In granular computing based on rough sets,neighborhood granularity based on neighborhood relationships is one representation of information granulation.The neighborhood granularity generated by different conditional attributes,and finer neighborhood granularity can result in a stronger differentiation of samples by the conditional attributes.However,previous research has mostly focused on the global perspective of neighborhood granularity generated by conditional attributes on the whole sample set,which may have limitations.For example,multiple conditional attributes may generate similar neighborhood granularities,making it difficult to differentiate the conditional attributes further.To fill the gap,this paper proposes the concept of local neighborhood granularity from a local perspective.Different from global neighborhood granularity,firstly the whole sample set is partitioned when calculating local neighborhood granularity and then the neighborhood granularity computed for each partition.Based on local neighborhood granularity,this paper proposes an attribute reduction algorithm that uses local neighborhood granularity as a criterion and remove conditional attributes having weak discriminative power for the samples when selecting conditional attributes.The proposed algorithm is compared with other algorithms using two different metric criteria in experiments performed on 12 UCI datasets.The experimental results show that the proposed algorithm has significant advantages in classifying accuracy and stability compared with the other algorithms.

关 键 词：属性约简 邻域粒度 邻域粗糙集 局部邻域粒度 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]