最大化独立有效分类信息率的属性选择

Attribute Selection via Maximizing Independent-and-Effective Classification Information Ratio

粗糙集中的属性选择有着十分重要的应用价值。现有的属性选择方法大多忽视了衡量待选属性所提供的分类信息和冗余信息,以及新增待选属性时已选属性所保留的分类信息三者之间的关联。因此,首先利用传统互信息,定义了有效分类信息率的属性重要性评估函数,并提出了一种基于有效分类信息率的属性选择方法。该属性选择方法可以有效地选择能提供大量有效分类信息同时携带较少冗余信息的待选属性。另外,考虑到新增待选属性对已选属性所保留的分类信息的影响,进一步提出了独立有效分类信息率的概念,并构造一种基于独立分类有效信息率的改进属性选择方法。该改进的属性选择方法能够有助于平衡属性的有效分类信息和冗余信息的关系,同时提高属性子集的整体识别能力。最后,从分类性能和统计学检验等方面分别与现有的属性选择方法进行了对比实验,实验结果表明了所提出的两种属性选择方法的有效性。

Attribute selection in rough set theory has wide practical application values.Most existing attribute selection approaches neglect the relationship among the classification information and redundant information brought by the candidate attribute,and the retained classification information provided by the selected attributes when selecting the candidate attribute.Therefore,the significant evaluation function of effective classification information ratio is defined for attribute selection,and an attribute selection approach via the effective classification information ratio is proposed further,which can effectively select the attributes that can provide lots of effective classification information and low redundant information.Besides,considering the influence of candidate attribute on the retained classification information provided by the selected attributes,another significant evaluation function of independent-and-effective classification information ratio is advanced,and an improved attribute selection approach is proposed,which can contribute to balancing the relationship between the effective classification information and redundant information of the attributes,and improving the overall recognition ability of the selected attribute subset.Finally,comparative experiments are conducted from the aspects of classification performance and statistical Bonferroni-Dunn test,and the experimental results illustrate that the proposed attribute selection approaches are effective.