Fuzzy bifocal disambiguation for partial multi-label learning

In partial multi-label learning (PML), each instance is associated with multiple candidate labels, but only a subset is the ground-truth label. Due to the ambiguous label information, PML is more challenging than traditional multi-label learning. Conventional PML mainly focuses on learning a desired feature space or label space for disambiguation, ignoring the tight correlation between two spaces. To this end, a novel Fuzzy Bifocal Disambiguation for Partial Multi-Label Learning (FBD-PML) method is proposed by employing the fuzzy label confidence to connect label and feature spaces so that these double spaces alternatively refine the fuzzy label confidence for connecting themselves better. FBD-PML learns the predictive label in the label space while using the fuzzy label confidence learned from the feature space to improve the discriminative ability of the classifier. Moreover, the manifold embedding is used to preserve the structure consistency between the original data and the fuzzy label confidence, encouraging amore accurate structure for the predictive label. Extensive experiments are conducted on a number of different data sets with different evaluation metrics and consistently demonstrate the superiority of FBD-PML.