Optimal Iterative Discriminant Analysis In Kernel Space

Kernel trick is a powerful tool being used for solving complex pattern classification problem. As long as a linear feature extraction algorithm can be expressed exclusively by dot-products, it can be extended to non-linear version by combining kernel method. In this paper, we present such an improved iterative algorithm used for linear discriminant analysis. By mapping data onto high dimensional feature space suing kernel function, we make data linearly separable and run iterative LDA there. Experiments with minimum distance classifier and nearest neighbor classifier show that our improved algorithm has a better performance than traditional Fisher discriminant and standard iterative LDA.