Low-Rank Multilabel Learning Based on Nonlinear Mapping

The relationship between features and labels, as well as the correlation between labels are two important factors that affect multilabel learning. Many existing studies assume that there is a linear mapping between sample space and label space. However, the linearity assumption often does not hold in practical applications, especially when characterizing the relationship between two high-dimensional spaces. In this article, a nonlinear mapping is used to describe the relationship between features and labels by mapping the original labels to a low-rank latent label space. With the help of nonlinear mapping, more compact and accurate label information can be mined, effectively overcoming the defect that linear models cannot fit real labels. In particular, the label space is firstly decomposed into a low-rank latent space to reduce the dimension of labels, and then a nonlinear mapping from the sample space to the low-rank latent space is constructed. Considering that the predicted values of similar labels should be closer, a comprehensive manifold regularization term is introduced to learn local and global label-correlation. The proposed multilabel learning method can effectively solve the problem of missing multilabel learning. The experimental results on ten multilabel databases show that the proposed method has good robustness and is superior to the most advanced multilabel learning methods.