Dimensionality reduction by regularized least squares weighted discriminant projection
Tian Qiang, Zhaolei Liu, Qiang Huang, Zhi Zhang, Zhanye Chen, Hanwen Chen · 2021 CIE International Conference on Radar (Radar) · 2021
In this paper, a new dimensionality reduction method, named regularized least squares weighted discriminant projection (RLSWDP), is developed to process high-dimensional signal which widely exists in real-world applications. In RLSWDP, all the representation coefficients are involved for samples reconstruction to improve reconstruction accuracy. Besides, the representation coefficients of each training sample are further used for weighting the reconstructed between-class scatter and within-class scatter which can preserve structural similarity between the samples resulting in better discrimination in projection space. Thirdly, a scale factor is introduced into the within-class weighting matrix to enhance the within-class compactness. Using the idea of maximum margin criterion, the optimal subspace projection matrix is obtained by eigenvalue decomposition. Experimental results on three face datasets demonstrate that the proposed method has promising performance compared to some popular dimensionality reduction methods.