Sparse representation based spectral regression

Spectral regression is a newly proposed method for dimensionality reduction, which is also based on graph embedding but less time and space consuming. However, like many methods based on neighborhood graphs, it still focuses on local manifold smoothness and ignores discriminative information among samples. In this paper, instead of making use of a neighborhood graph, we take advantage of a global graph defined by the coefficients of sparse representation and propose a method called Sparse Representation based Spectral Regression (SpSR) on this graph. This graph is data-driven, discriminative and robust to noise features. Experimental results on facial images feature extraction tasks demonstrate these advantages.