In partial label learning, each training object is assigned a valid label and pseudo-labels, and a multi-class classifier is derived with inaccurate supervision. However, ambiguous labeling information adversely affects the performance of the classifier. Partial label feature selection has been shown efficiently improve the generalization performance of classifiers. Traditional manifold learning can employ intrinsic geometric information to identify discriminative features, while it is challenging due to the noisy manifold caused by pseudo-labels. Consequently, this paper proposes an embedding partial label feature selection based on noisy manifold and label distribution, which exploits feature dependency, label correlation, and instance relevance. Specifically, a linear regression function projects the feature space to the low-dimensional manifold space, which can avoid the influence of pseudo-labels affected by direct projection to the label space. The feature dependency and label correlation are obtained by manifold regularization in the feature and label space to reflect the feature significance. During optimization, instance similarity constraints variable iteration. Label distribution obtained through feature significance and instance relevance guides label space updates and reduces the impact of noise in the manifold. The effectiveness and robustness of the proposed algorithm are corroborated through experiments with three classifiers and five comparison methods on twelve datasets.