Abstract: The goal of partial multi-label learning is to train a noise-robust classifier that can accurately assign labels to unknown instances, even when the candidate label set is only partially valid. Currently, most existing approaches rely on the label correlation assumption: the correlations between label categories maintain consistency across different datasets. However, the presence of noise makes the obtained prior label correlation unreliable. To tackle this problem, a novel partial multi-label learning approach with fuzzy weakly supervised label correlation refinement is proposed. First, a fuzzy framework is established to learn fuzzy label membership through clustering analysis, which can be viewed as fuzzy weakly supervised label information since it measures the distances between instances and class prototypes. This research first proposes to replace the prior label correlation with fuzzy label correlation measured by fuzzy label membership while preserving certain global structures of the original label space. By encouraging consistency correspondence of sample manifolds in label confidence, this new label correlation, together with sample similarity, is jointly employed to learn more precise label confidence. Finally, the learned label confidence is exploited to train a kernelized linear classifier for unseen instances. This proposed method is solved by using an effective iterative optimization algorithm. Extensive experiments demonstrate that the proposed model exhibits superior performance compared with other advanced partial multi-label learning methods.