<p>Partial multi-label learning (PML) addresses the problem where some instances are associated with a candidate label set containing noisy labels. While most existing methods primarily focus on using label information to handle label noises, they often overlook the rich information embedded in instance features. In order to fully utilize the label and feature information, a novel label representation and feature-aware learning method is proposed. Specifically, the quality of labeling is improved by learning the label representation. It is also combined with feature-awareness to optimize the classifier and ensure that the predicted label information is consistent with the label representation. The model strengthens the label structure through instance and feature manifold learning to make the final classification results more robust. Extensive experiments conducted on several datasets from various domains demonstrate the effectiveness and superiority of the proposed approach compared to state-of-the-art methods across multiple evaluation metrics.</p>