基于模糊邻域相对依赖互信息的特征选择方法

Feature Selection Method Based on Fuzzy Neighborhood Relative Dependency Mutual Information

模糊粗糙集作为处理不确定性信息的有效工具,已广泛应用于特征选择中。然而当数据分布密度差别较大时,传统模糊粗糙近似不能有效度量样本的隶属度,且大多特征评价函数仅从代数或信息单一视角构造。针对以上问题,提出了一种基于模糊邻域相对依赖互信息的特征选择方法。首先,为克服传统模糊粗糙近似对数据分布敏感的缺陷,引入相对距离计算模糊相似关系,同时考虑模糊邻域粒度结构,提出了模糊邻域相对依赖度,从代数观度量数据的不确定性。然后,基于相对粒度结构提出了模糊邻域相对互信息,并与模糊邻域相对依赖度结合构造出一种新的特征评价函数——模糊邻域相对依赖互信息,将代数观和信息观结合进行特征评价。最后,设计了一种基于模糊邻域相对依赖互信息的特征选择算法(FNRDI)。通过与其他算法在9个公共数据集上进行实验对比分析,结果表明所提算法可有效消除冗余特征且提高数据的分类精度。

As an effective tool to deal with uncertain information,fuzzy rough set has been widely used in feature selection.However,when the data distribution density is quite different,traditional fuzzy rough approximation can’t effectively measure the membership degree of samples,and most feature evaluation functions are constructed only from a single perspective of algebra or information.To solve these problems,a feature selection method based on fuzzy neighborhood relative dependency mutual information is proposed.First,in order to overcome the defect that traditional fuzzy rough approximation is sensitive to data distribution,the relative distance is introduced to calculate the fuzzy similarity relation,meanwhile,considering the fuzzy neighborhood granular structure,the fuzzy neighborhood relative dependency degree is proposed to measure the uncertainty of data from algebraic aspect.Then,the fuzzy neighborhood relative mutual information based on the relative granular structure is proposed,and combined with the fuzzy neighborhood relative dependency degree to construct the fuzzy neighborhood relative dependency mutual information as a new feature evaluation function,to evaluate the features from algebra and information view.Finally,a feature selection algorithm based on fuzzy neighborhood relative dependency mutual information is designed.Through experimental comparative analysis with other algorithms on 9 public datasets,show that the proposed algorithm can effectively eliminate redundant features and improve the classification accuracy of data.