Noisy feature decomposition-based multi-label learning with missing labels

In recent years, multi-label learning with missing labels (MLML) has become a popular topic. The major challenge for MLML is enhancing the performance of classifiers in the presence of missing labels. Most existing algorithms focus on recovering missing labels using label correlations. However, incomplete label correlations in the early stages of recovery may adversely affect the results. To address this problem, we focus on the original task of finding the mapping between labels and features and propose a Noisy Feature Decomposition-based Multi-label learning with Missing Labels (NFDMML) method. Specifically, the label information is assumed to be integral, and the features corresponding to missing labels are defined as noisy features. Not recovering the missing labels, we reduce the interference of noisy features in the classifications. Accordingly, the MLML problem is converted into a feature decomposition problem. Based on label correlation, a low-rank relationship is used to eliminate the features caused by missing labels, and reverse mapping is employed to preserve the features corresponding to the relevant labels. We conduct detailed experiments on multiple datasets, and the results clearly demonstrate that the proposed method achieves competitive performance over other algorithms.