基于类属属性约简的多标记学习  

Multi-label learning with label-specific feature reduction

作　　者：徐苏平[1] 杨习贝[1] 祁云嵩[1] 

机构地区：[1]江苏科技大学计算机科学与工程学院,江苏镇江212003

出　　处：《计算机应用》2015年第11期3218-3221,3226,共5页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61471182;61100116;61305058);江苏省自然科学基金资助项目(BK2012700;BK20130471);中国博士后科学基金资助项目(2014M550293)

摘　　要：在多标记学习中,由于不同的标记可能会带有自身的一些特性,所以目前已经出现了基于标记类属属性的多标记学习算法LIFT。然而,类属属性的构建可能会增加属性向量的维度,致使属性空间存在冗余信息。为此,借助模糊粗糙集提出了一种能够进行类属属性约简的多标记学习算法FRS-LIFT,其包含4个步骤:类属属性构建、属性维度约简、分类模型训练和未知样本预测。在5个多标记数据集上的实验结果表明,该算法与LIFT算法相比,不仅能够降低类属属性维数,而且在5种多标记评价指标上均具有较好的实验效果。In multi-label learning, since different labels may have their own characteristics, multi-label learning approach with label-specific features named LIFT has been proposed. However, the construction of label-specific features may increase the dimension of feature vector, which brings some redundant information in feature space. To solve this problem, a multi- label learning approach named FRS-LIFT was presented, which can implement label-specific feature reduction by fuzzy rough set. FRS-LIFT contains four steps： construction of label-specific features, reduction of feature dimensionality, training of classification models and prediction of unknown samples. The experimental results on 5 multi-label datasets show that, compared with LIFT, the proposed method can not only reduce the dimension of label-specific features, but also achieve satisfactory performances in 5 evaluation metrics.

关 键 词：属性约简 模糊粗糙集 类属属性 多标记学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]