Abstract: Attribute reduction is the common data preprocessing method in such areas as machine learning. In the attribute reduction algorithm based on rough set theory, most of which based on single method to evaluate the importance of attribute. In order to achieve more accurate measurement from multiple perspectives for attribute, the concept of attribute dependency measurement is defined in the existed fuzzy neighborhood rough set model. Then, according to the conception of knowledge granularity in granular computing theory, the fuzzy neighborhood granularity measurement is proposed under the model of fuzzy neighborhood rough set. Because of the attribute dependency and knowledge granularity are representing a different respective of method of attribute evaluation, the two measurement methods are combined as the method of attribute importance evaluation for information system. Finally, a heuristic attribute reduction algorithm is given. The experimental results show that the proposed algorithm has better attribute reduction performance.

Key words: attribute reduction, fuzzy neighborhood rough set, dependency, knowledge granularity, fuzzy neighborhood granularity

摘要： 属性约简是机器学习等领域中常用的数据预处理方法。在基于粗糙集理论的属性约简算法中，大多是根据单一的方法来度量属性重要度。为了从多角度对属性达到更为优越的评估效果，首先在已有的模糊邻域粗糙集模型中定义属性依赖度度量，然后根据粒计算理论中知识粒度的概念，在模糊邻域粗糙集模型下提出了模糊邻域粒度度量。由于属性依赖度和知识粒度代表了不同视角的属性评估方法，因此将这两种方法结合起来用于信息系统的属性重要度评估，最后给出一种启发式属性约简算法。实验结果表明，所提出的算法具有较好的属性约简性能。

关键词: 属性约简, 模糊邻域粗糙集, 依赖度, 知识粒度, 模糊邻域粒度