Abstract: In traditional heuristic algorithm to compute reduction, one and only one measurement is used to construct the constraint. Therefore, the derived reduction may not always meet the constraints based on multi-measurement. Moreover, most of the definitions of attribute reductions are given by considering all decision classes, while ignoring the changes in the metrics corresponding to different decision categories before and after reduction. To solve these problems, an attribute reduction strategy for local multi-constraint is proposed, which aims to make each decision class meet the constraints based on multi-measurement. By using neighborhood rough set, the traditional strategy is compared with the local multi-constraint strategy in computing reductions on UCI data sets. The approximate quality and conditional entropy are selected as the measurements used in multi-constraint. The experimental results show that the local multi-constraint based attribute reduction can significantly reduce the conditional entropy while it meets the constraint of the approximate quality. The effectiveness of local multi-constraint strategy is then displayed.

Key words: approximate quality, attribute reduction, conditional entropy, neighborhood rough set

摘要： 传统求解约简的启发式算法采用单一的度量指标作为约束条件，但这一策略并不能保证约简满足多重度量指标下约束需求。除此之外，绝大多数的约简定义是建立在考虑所有决策类基础上的，而忽视了不同决策类别所对应的度量指标在约简前后的变化情况。针对这些问题，提出了一种面向局部多约束的属性约简策略，其目的是使得每个决策类别都能够满足多重度量指标下的约束条件。借助邻域粗糙集模型，在UCI数据集上将传统约简策略与局部多约束约简策略进行了对比分析，将近似质量与条件熵作为多重约束中的度量指标，实验结果表明局部多约束约简能够在近似质量满足约束条件的前提下，显著降低条件熵，从而展现了局部多约束约简的有效性。

关键词: 近似质量, 属性约简, 条件熵, 邻域粗糙集