Learning common and label-specific features for multi-Label classification with correlation information

作者：Li, Junlong; Li, Peipei^*; Hu, Xuegang; Yu, Kui

来源：Pattern Recognition, 2022, 121: 108259.

DOI：10.1016/j.patcog.2021.108259

• Multi-label classification
• Label-specific features
• Common features
• Instance correlation

In multi-label classification, many existing works only pay attention to the label-specific features and label correlation while they ignore the common features and instance correlation, which are also essential for building a competitive classifier. Besides, existing works usually depend on the assumption that they tend to have the similar label-specific features if two labels are correlated. However, this assumption cannot always hold in some cases. Therefore, in this paper, we propose a new approach of learning common and label-specific features for multi-label classification using the correlation information from labels and instances. First, we introduce l(2,1)-norm and l(1)-norm regularizers to learn common and label-specific features simultaneously. Second, we use a regularizer to constrain label correlations on label outputs instead of coefficient matrix. Finally, instance correlations are also considered through the k-nearest neighbor mechanism. Comprehensive experiments manifest the superiority of our proposed approach against other well-established multi-label learning algorithms for label-specific features.