Generalization of linear discriminant analysis using LpLp-norm

In this paper, the linear discriminant analysis (LDA) is generalized by using an LpLp-norm optimization technique. Although conventional LDA based on the L2L2-norm has been successful for many classification problems, performances can degrade with the presence of outliers. The effect of outliers which is exacerbated by the use of the L2L2-norm can cause this phenomenon. To cope with this problem, we propose an LDA based on the LpLp-norm optimization technique (LDA-LpLp), which is robust to outliers. Arbitrary values of p can be used in this scheme. The experimental results show that the proposed method achieves high recognition rate for many datasets. The reason for the performance improvements is also analyzed.

► LDA is generalized to use Lp-norm in both numerator and denominator. ► Steepest gradient method is used for optimization. ► The effect of outliers are analysed.