Low-Rank Correlation Analysis for Discriminative Subspace Learning

Linear dimensionality reduction is a commonly used technique to solve the curse of dimensionality problem in pattern recognition. However, learning a discriminative subspace without label is still a challenging problem, especially when the high-dimensional data is grossly corrupted. To address this problem, we propose an unsupervised dimensionality reduction method called Low-Rank Correlation Analysis (LRCA). The proposed model integrates the low-rank representation and the linear embedding together with a seamless formulation. As such, the robustness and discriminative ability of the learned subspace can be effectively promoted together. An iterative algorithm equipped with alternating direction method of multiplier (ADMM) and eigendecomposition is designed to solve the optimization problem. Experiments show that our method is more discriminative and robust than some existing methods.