Abstract: Traditional criteria of attribute reduction for neighborhood decision error rate is designed for improving overall classification accuracy, it does not take the variation of accuracy of each class into consideration when reduction finding is executed. From this point of view, the concepts of local neighborhood decision error rate and local attribute reduction are introduced for improving the classification accuracy of single class. Furthermore, a heuristic algorithm to compute local neighborhood decision error rate based reduction is presented. The experimental results on 8 UCI data sets show that the local reduction can not only improve the classification accuracy of single class, but also overcome the limitation of accuracy’s decreasing for single class, which may be caused by global reduction.

Key words: attribute reduction, global reduction, heuristic algorithm, local reduction, neighborhood rough set

摘要： 传统基于邻域决策错误率的属性约简准则是针对总体分类精度的提升而设计的，未能展现因约简而引起的各类别精度变化情况。针对这一问题，引入局部邻域决策错误率以及局部属性约简的概念，其目的是提升单个类别的分类精度。在此基础上，进一步给出了求解局部邻域决策错误率约简的启发式算法。在8个UCI数据集上的实验结果表明，局部约简不仅是提高各个类别精度的有效技术手段，而且也解决了因全局约简所引起的局部分类精度下降问题。

关键词: 属性约简, 全局约简, 启发式算法, 局部约简, 邻域粗糙集