Label Disambiguation-Based Feature Selection for Partial Multi-label Learning

Partial multi-label learning (PML) addresses the issue of training a multi-label predictor in the context of inaccurate supervision. Objects in PML are relevant to multiple semantics, but only a subset of them are valid. Besides false positive labels that mislead the learning procedure, high dimensionality also acts as a stumbling block for boosting PML. In this paper, a two-stage label disambiguation-based feature selection method, LDFS-PML, is presented for partial multi-label learning. At first, to avoid false positive labels from misleading the feature selection, a label disambiguation technique is devised based on the granular ball, which is the first attempt at multi-label disambiguation from the perspective of cognition computing. By using the label disambiguation technique, label-specific information concealed in the distribution of data is captured, which is conducive to estimating the confidence of candidate labels. In the second stage of LDFS-PML, a feature selection algorithm is proposed which utilizes labeling confidence and simultaneously incorporates cognition computing from both global and local perspectives. Experiments are conducted on various PML datasets, and the superiority of the proposed LDFS-PML is demonstrated.