Compute orthogonal discriminant vectors efficiently for dimension reduction

Fisher discriminant analysis is a popular technique for dimension reduction and feature extraction. However, the discriminant vectors in the naïve Fisher discriminant analysis are often correlated with each other. We propose an efficient method for computing orthogonal discriminant analysis vectors for dimension reduction in Fisher discriminant analysis. In the proposed method, while the first discriminant vector is worked out in the sample space, the others are worked out by maximizing a Fisher criterion defined in a transformed space which is the null space of the previously obtained discriminant vectors. We also propose two algorithms to implement the model. The experimental results show that the proposed method is effective and efficient.