Dimensionality Reduction Using Discriminant Collaborative Locality Preserving Projections

Abstract

In this paper, we propose an effective dimensionality reduction algorithm named Discriminant Collaborative Locality Preserving Projections (DCLPP), which takes advantage of manifold learning and collaborative representation. Firstly, two adjacency graphs of the input data are adaptively constructed by an l2-optimization problem to model discriminant manifold structure. The adjacency graphs characterize the important properties such as the intra-class compactness and the inter-class separability. Next, based on collaborative representation reconstruction weights, both intra-class collaborative representation scatter and inter-class collaborative representation scatter can be calculated. Then, motivated by MMC, DCLPP can obtain optimal projection directions which could maximize the between-class scatter and minimize the within-class compactness. DCLPP naturally avoids the small sample size problem. Finally, after dimension reduction and data projection by DCLPP, the NN classifier is employed for classification. To evaluate the performance of DCLPP, we compare it with the most existing DR methods such as CRP and DSNPE on publicly available face databases and COIL-20 database. The experimental results demonstrate that DCLPP is feasible and effective.