Classifier enhancement based on credible sample selection for partial multi-label learning

Partial multi-label learning (PML) is a weakly supervised framework where each training sample is associated with several candidate labels, which include noisy labels. The main goal is to overcome the noise interference and achieve a well-trained classifier. Given that the sample features contain redundancy and the sample labels include noise, these factors can introduce interference during classifier training. Therefore, we aim to construct the sample set that prioritizes those with less noise, higher representativeness and confidence to improve the effectiveness of the model. To achieve this, we propose a new PML approach with classifier enhancement based on credible sample selection, called PML-CECS. Specifically, this paper first projects the feature space and label space into the subset space, enhancing the consistency of representation within the subset space by sharing projection information during this process. Then orthogonalization is applied to the subset space to reduce noise and redundant correlations, thereby improving the representativeness and reliability of the data. Next, the manifold structure reinforces the instance-level consistency between features and labels within the subset space. And leveraging the subset samples as new learning information further enhances the classifier’s performance. Finally, to mitigate erroneous correlations arising from noise interference, pseudo-labels are introduced and integrated into the model training. Extensive experiments have validated the feasibility of this approach.