Dimensionality Reduction Based on Low Rank Representation

Dimensionality Reduction is a common way to solve the problem of ‘curse of dimensions’, especially for image processing. Among all these methods, the linear methods are believed to have better performance in actual databases. This paper proposes a novel unsupervised linear dimensionality reduction method that based on low rank representation which aims at finding the subspace structure of the original data sets. This method named LRRDR tries to preserve the subspace structure of the original data therefore is better than the global dimensionality reduction methods like PCA. The experiments compare PCA, NPE, LRRDR and SPP and the results show that LRRDR outperforms the other methods.