基于特征关联的多标记谱特征选择算法  

Multi-label Spectral Feature Selection Based on Feature Correlation

作　　者：胡敏杰[1] 杨红和[2] 傅为[1] 郑荔平[1] 

机构地区：[1]闽南师范大学计算机学院,福建漳州363000 [2]闽南师范大学教务处,福建漳州363000

出　　处：《闽南师范大学学报（自然科学版）》2017年第2期12-22,共11页Journal of Minnan Normal University：Natural Science

摘　　要：在多标记学习中,一个样本不仅同时具有多个标记,而且还由高维特征描述,解决特征间的相关性和冗余性以应对维数灾难问题是多标记学习前的一个重要步骤.文中将谱图理论与多标记数据的特征选择相结合,提出基于特征关联的多标记谱特征选择算法.该算法假设各标记之间独立,首先对各单个标记考虑特征间的相关性和冗余性构建目标函数生成与该标记相关的一组特征排序;然后对各标记产生的排序进行融合,最后得到一组合理的特征序列.在4个数据集和5个评价指标上的实验表明,文中算法优于当前流行的多标记特征选择算法.In the multi-label learning a sample may be assigned with more than one decision label, and also described with high-dimensional vectors, solving the correlation and redundancy among features to conquer the curse of dimensionality is an important step in multi-label learning before. In this paper, the spectral theory is applied to the feature selection of multi-label data. The algorithm assumes that each independent label. First, it considers the correlation and redundancy between features based on different labels to construct objective function ,which can generate a set of features under each label; Then, it fuses all individual feature rank lists and gets a reasonable feature sequence. Experiments on four datasets and five evaluation criteria show that the proposed algorithm is superior to several state -of -the -art multi -label feature selection algorithms.

关 键 词：多标记分类 特征选择 特征关联 谱图理论 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]