Local Weighted Semi-supervised Discriminant Analysis for Dimensionality Reduction

Abstract

In this paper, we present a novel weighted version of semi-supervised discriminant analysis method by assigning weights to each labeled samples. The proposed within-class weight can detect the outliers and between-class weight can discover the support points in boundaries between different classes. In addition, our proposed method is robust to diversedensity classes and imbalanced boundaries. For highdimensional dataset, our method can find a nice lowdimensional projection to preserve the discriminative information and manifold structure embedded in both labeled and unlabeled samples. It can also be easily kernelized to form a nonlinear method and do semi-supervised induction.The experiments show that our method can achieve very promising classification accuracies than other methods.

Multiprojection Local Discriminant Analysis

“…By maximizing the ratio of between-class distance to within-class distance in the subspace, LDA maximizes the separability of the projected instacnes. Specifically tailored, LDA can also serve as a semi-supervised algorithm [6,7] for incompletely labelled data. Many stronger variants of LDA have also been constructed from different perspectives.…”