基于层次类别邻域粗糙集的在线流特征选择算法

Online streaming feature selection method based on hierarchical class neighborhood rough set

在开放动态环境中,在线流特征选择是降低特征空间维度的有效方法 .现有的在线流特征选择算法能够有效地选择一个较优的特征子集,然而,这些算法忽略了类别中可能存在的层次结构.基于此,提出基于层次类别邻域粗糙集的在线流特征选择算法:首先,在邻域粗糙集中引入层次最近异类的邻域关系,避免邻域粒度的选择,借助层次结构计算特征对标记的层次依赖度,推广邻域粗糙集模型以适应层次类别数据;其次,基于层次依赖度提出三个在线特征评价函数,设计了在线相关选择、在线重要度计算和在线冗余更新的层次特征选择框架;最后,在六个层次类别数据集和八个扁平单标记数据集上的实验表明,提出的算法优于现有最先进的在线流特征选择算法.

Online streaming feature selection is an effective approach to reduce feature space dimension in open and dynamic environment. Existing streaming feature selection methods can effectively get an optimal feature subset. However,these methods generally ignore a real-world scenario,i. e.,the hierarchical structure of the class. To address this problem,we propose an online streaming feature selection method based on hierarchical class neighborhood rough set. Firstly,we introduce a new neighborhood relation to avoid granularity selection. Then,the hierarchical structure is leveraged to calculate hierarchical dependence,which generalize neighborhood rough set to fit hierarchical classification learning. Three online feature subset evaluation functions based on hierarchical dependence are defined. We design an online hierarchical label streaming feature selection framework including online relevance selection,online importance calculation,and online redundant updation. Finally,empirical experiments are carried out on six hierarchical class and eight single label datasets,showing that the proposed algorithm outperforms other state-of-the-art online streaming feature selection methods.