Abstract:
|
Regularization is a key component in high dimensional data analyses. We'll present some related work on the regularization of Fisher's linear discriminant analysis (LDA) in its applications to high dimensional data discrimination. Some considerations when categorizing different methods or choosing one method over the others are discussed. Fundamental differences are in the choice of the criterion, that is whether a method aims to modify the original LDA solution, or tries to solve Fisher's original optimization problem, or tries to find a solution suitable for high dimensional data to an equation such as generalized eigenvalue problem. An important consideration is whether a method is limited to the binary case or an extension to the multi-category cases is straightforward, while some methods directly start with multi-category problems. In practice the computation is a critical issue, which gives a method based on an efficient optimization algorithm a big advantage.
|