A Sparse Local Preserving Projection Method Based On Graph Embedding

Abstract

Dimensionality reduction plays an important role in pattern classification. In this paper, a robust unsupervised dimensionality reduction method termed robust sparse locality preserving projection with adaptive graph embedding is proposed. Specifically, the proposed method integrates the adaptive graph learning and projection learning into a framework, which can capture the intrinsic locality structure of data and in turn promotes the method to achieve the global optimal projection. To capture the global information of data, a variant PCA term is introduced, which can decrease the information loss during dimensionality reduction. Importantly, a row-sparsity constraint is imposed on the projection to select the most important features for dimensionality reduction, so as to improve the robustness of the proposed method to noises. Extensive experiments are performed on three representative face databases and an object database, which sufficiently validates the superiority of the proposed method in comparison with some state-of-the-art-methods.