Abstract: In the multi-label classification problems, different labels corresponding with the same instance may have some correlations, and the label correlation can provide some important information. Some multi-label learning methods are presented to combine the label correlation, but they are built on the original data which are high-dimensional and noisy, then they cannot obtain the satisfied results. This paper proposes a common subspace based multi-label learning method to extract high-level information from the original feature space by making use of the label information. Based on the feature mapping function, the original high-dimensional data can be effectively represented in the common low-dimensional subspace. Meanwhile, the proposed method also learns the mapping function from the label space to the common high-level feature space, then the label information of instances can be obtained with the new representation in the subspace. The experimental results on five real-world data sets show that the model can effectively improve the performance of multi-label classification.

Key words: multi-label learning, common subspace, label correlation

摘要： 在多标记问题中，一个样本对应的多个类别之间经常会存在一定的相关性，这些相关性可以为多标记分类提供有用的信息。已有的多标记学习对于类别之间的相关性研究是建立在原始数据上的，然而原始数据往往是高维且含有噪声的，使得已有学习方法无法达到满意的效果。提出了一种基于共享子空间的多标记学习方法。该方法可以在类别信息的指导下，学到从原始特征空间到高层共享空间的映射函数，从而可以把原始的高维数据映射到一个低维空间中。同时也学到一个从类别空间到高层空间的映射函数，使得数据进行低维的重新表示后，可以直接对应到类别信息。在5个实际的数据集合上进行了测试，实验结果表明该模型可以有效地提高多标记数据的分类性能。

关键词: 多标记学习, 共享子空间, 类别相关性