Abstract: Low-Rank Representation (LRR) allows for the representation of each data point as a linear combination of bases, making it a promising approach for capturing underlying low-dimensional structures. However, most LRR methods use the original dataset as a dictionary, which cannot reveal the true segmentation of the data. In this paper, we propose an unsupervised projection learning method, called Manifold Projection Learning via Low-Rank Representation with Subspace Dictionary (MPL-LRRSD). MPL-LRRSD learns an optimal subspace as the dictionary for the LRR problem instead of using the original dataset. The original data can be well recovered by low-rank matrix using minimal bases. Meanwhile, by imposing row sparse constraint on the projection matrix, MPL-LRRSD not only selects discriminative features and eliminates the redundant features, but also makes the subspace learning well interpretable. Furthermore, we introduce manifold structure preserving constraint to preserve both original representation and distance information of the samples under projection. Extensive experiment results on various real-world datasets demonstrate the superiority over the state-of-the-art methods.