Multi-label classification is an appealing and challenging supervised learning problem, where multiple labels, rather than a single label, are associated with an unseen test instance. To remove possible noises in labels and features of high-dimensionality, multi-label dimension reduction has attracted more and more attentions in recent years. The existing methods usually suffer from several problems, such as ignoring label outliers and label correlations. In addition, most of them emphasize on conducting dimension reduction in an unsupervised or supervised way, therefore, unable to utilize the label information or a large amount of unlabeled data to improve the performance. In order to cope with these problems, we propose a novel method termed Robust sEmi-supervised multi-lAbel DimEnsion Reduction, shortly READER. From the viewpoint of empirical risk minimization, READER selects most discriminative features for all the labels in a semi-supervised way. Specifically, the ℓ2,1-norm induced loss function and regularization term make READER robust to the outliers in the data points. READER finds a feature subspace so as to keep originally neighbor instances close and embeds labels into a low-dimensional latent space nonlinearly. To optimize the objective function, an efficient algorithm is developed with convergence property. Extensive empirical studies on real-world datasets demonstrate the superior performance of the proposed method.