<p>The partial multi-label learning task assumes that instances in real-world scenarios often contain labels with ambiguity and discrimination. Label disambiguation is a key approach to solving the PML problem. However, existing label disambiguation PML methods based on label correlation knowledge are not sufficient in mining label structure relationships. On the one hand, most of the methods ignore the effect of noisy labels in the candidate labels on the label correlation modeling, which leads to the poor robustness of the models; on the other hand, these methods fail to thoroughly explore the structural properties of the label space at the label level, which results in the learned label relations being inaccurate or even having errors. We propose a novel Label-Driven Global high-rank and Local low-rank partial multi-label learning method named LDGL to address the issues mentioned above. Specifically, firstly, the influence of noisy labels is effectively mitigated by explicit global high-rank and local low-rank structural learning in the candidate label space, combined with the knowledge of label correlation. Secondly, we impose low-rank constraints on the weight coefficient matrices to motivate the sharing of common information among features and improve the capacity of the model for generalization. Finally, the label smoothness and consistency assumptions are utilized to evaluate the confidence of ground-truth labels, and the label structure learning and label disambiguation are implemented in a unified learning framework, which learns from each other to promote each other. We evaluated it on 7 synthetic and 3 real-world PML datasets, using 7 state-of-the-art PML algorithms as baselines and 5 evaluation metrics. Out of 155 experimental scenarios, the proposed method outperforms the baselines by about 79.4%, demonstrating its robustness and effectiveness in handling PML tasks.</p>