Multi-label learning for label-specific features using correlation information with missing label

Multi-label learning tasks often encounter challenges in acquiring complete labels. Numerous algorithms address this issue by leveraging label correlation or instance correlation to recover missing labels. However, these approaches may overlook the intrinsic relationship between instances and labels, which could also be advantageous for recovering missing labels. This study introduces a competitive multi-label learning algorithm called Label-Feature Learning with Missing Label Correlation Information (LFLI) to tackle this problem. Firstly, the missing label matrix is completed using label correlations guided by reinforcement from instance correlations. Subsequently, label-specific features are selected, and finally, the object model is constructed by integrating the correlation information (including label correlations and instance correlations), missing labels, and label-specific features. Experimental results demonstrate that LFLI exhibits strong competitiveness compared to several state-of-the-art algorithms across four evaluation metrics on benchmark multi-label classification datasets.