Multi-label feature selection is an efficient technique to alleviate the high dimensionality for multi-label learning. Existing multi-label feature selection methods based on information theory either deal with labels individually or treat all label relationships as redundancy. However, two important and being ignored issues are the different effects of label relationships and the dynamic changes of label relationships in measuring different candidate features. To address these issues, we first distinguish three types of label relationships: label independence, label redundancy and label supplementation. Second, we consider the changes of label relationships based on different features. By analyzing the differences and the changes of label relationships, two new methods named LSMFS and MLSMFS are proposed, which extracts all supplementary information and the maximum supplementary information of features for each label from other labels, respectively. Finally, experiments on fifteen benchmark multi-label data sets demonstrate the effectiveness of the proposed methods against nine other methods.

多标签特征选择是缓解多标签学习高维的有效技术。现有的基于信息论的多标签特征选择方法要么单独处理标签，要么将所有标签关系视为冗余。然而，两个重要且被忽视的问题是标签关系的不同影响和标签关系在测量不同候选特征时的动态变化。为了解决这些问题，我们首先区分三种类型的标签关系：标签独立性、标签冗余和标签补充。其次，我们考虑基于不同特征的标签关系的变化。通过分析标签关系的差异和变化，提出了两种新方法LSMFS和MLSMFS，它分别从其他标签中提取每个标签的所有补充信息和特征的最大补充信息。最后，在 15 个基准多标签数据集上的实验证明了所提出的方法相对于其他 9 种方法的有效性。