关键词: 粗糙集, 差别矩阵, 最小属性约简, 划分粒度, 分解

Abstract: Many existing algorithms of attribute reduction begin at constructing decision table’s discernibility matrix，then convert non-empty objects’ conjunctive normal form into minimal disjunctive normal form.It is important how to get a reduction more efficiently.This paper points out that the minimum attribute reduction algorithm is imperfect in some respect，and an improved algorithm for the minimum attribute reduction based on U/{a} partition is proposed.By regarding the significance of attributes defined from the viewpoint of partition granularity as heuristic information，and introducing the heuristic information into U/{a} partition which translates attribute reduction problem in macrocosm into attribute reduction problem in subdomain.Theoretical analysis and example show that this algorithm is effective.

Key words: rough set, discernibility maxtrix, mininum attribute reduction, partition granularity, decomposition