Partial multi-label feature selection via subspace optimization

Feature selection is an effective way to improve the model learning performance while being a challenging problem in the Partial Multi-label Learning (PML). Different from multi-label learning, PML is closer to reality, which means the annotators may assign irrelevant labels at least one in the set of candidate labels for each instance. However, previous PML researches only design the model on label disambiguation but ignore to consider the problem as point of finding discriminative features in the noisy scenario. In this paper, a novel feature selection method named PML-FSSO is proposed which utilizes the theory of linear weighted to jointly consider the label subspace and feature subspace. The influence of noisy labels is eliminated by decomposing the label matrix into the low-dimensional space, and we call the low-dimensional space as the label subspace. Then, the feature matrix is not only designed to be the denoising matrix as the feature subspace, but also to steer the direction for label disambiguation. Finally, the coherent subspace is constructed through the shared coefficient weight matrix which is used in the linear weighted method. Extensive experiments clearly validate the effectiveness of PML-FSSO in improving the state-of-the-art feature selection methods in the field of partial multi-label learning.