Abstract:

Multi-label manifold learning (ML$^2$) constructs label manifolds based on feature manifolds and converts logical into numeric labels. This can better reflect the correlations between labels and improve classification performance. However, similar to most methods, ML$^2$ is based on all features and ignores different discriminabilities when different features are used to classify different labels. Therefore, a method we call label-specific feature-based multi-label manifold learning (LSF-ML$^2$) is proposed. First, the labels are used to optimise the feature importance matrix, which can determine the subset of label-specific features. Then, the feature manifold of the subset is mapped to the label space so that the logical labels can be converted into numeric labels. Finally, a multi-output regression is applied for classification. Experimental results show that the proposed method outperforms several existing multi-label classification methods.

Key words: multi-label learning, classification, label-specific feature, manifold learning, label correlation