决策代价约简求解中的交叉验证策略

Cross-validation strategy in attribute reduction based on decision cost

属性约简是粗糙集理论中的核心问题,其目的是剔除冗余属性以找到具有较好泛化能力的属性子集.在决策粗糙集理论中,决策代价经常被作为属性约简的约束条件.但值得注意的是,虽然基于决策代价的约简求解算法可以有效地降低训练样本集上的总决策代价,但其往往忽视了测试样本集上的总决策代价.为解决这一问题,利用交叉验证的基本思想,设计了以决策代价为约束条件的一种新的属性约简求解算法.在八个UCI数据集上的实验结果表明,相较于传统基于决策代价的约简求解算法,所提算法不仅能有效地降低训练集合和测试集合的总决策代价,而且找出的属性子集亦可以带来更好的分类性能。

Attribute reduction is a core problem in rough set theory,with its purpose of getting rid of redundant attributes to obtain a reduct with better generalized performance. In decision-theoretic rough set, the decision cost is frequently regarded as constraint of attribute reduction. However, it is worthy to notice that although the reduct obtained by the algorithm based on the consideration of decision cost can effectively reduce the decision cost of the training set, it may fails to effectively reduce the decision cost of the test set. To solve such problem,a new algorithm, based on the method of cross - validation, is designed through using the decision cost as constraint. The experimental result over eight UCI data sets show that compared with the traditional algorithm based on decision cost, the proposed algorithm not only reduces the decision cost of training set and the test set,but also brings better classification performance.