Consider other classifiers

[ramhiser]

• Add rda from the klaR package
• Drop Pang's classifier?
• Add kNN and SVM?

[ramhiser]

Here are some comments from AoAS reviewers regarding this matter.

First Reviewer

Major Comments

1. Comparing their regularized discriminant analysis method with those that assume independence (e.g. Dudoit, Pang, Tong, and Guo) is not a fair comparison. Classifiers that do not have this assumption should be compared.
2. Elastic net, k-nearest neighbor, and support vector machine are also excellent classifiers, these should be compared.

Second Reviewer

I have major concerns about the real data analysis. As far as I can tell, no comparisons are made to other approaches that use a compromise between LDA & QDA: for instance, equation (16c) in Friedman (1989), or the technique of http://arxiv.org/abs/1111.1687 . Instead, the comparisons are all to methods that are essentially doing either LDA or diagonal LDA, and so this is an apples-to-oranges comparison. To further confuse the comparison, the competitors perform built-in feature selection, but the proposed approach does not -- and yet pre-screening is done for all techniques.

A simulation study indicating, in an objective way, the domain on which this approach will outperform competitors (e.g. other competitors that marry LDA and QDA, such as those mentioned earlier) is missing.

Furthermore, the 1st full paragraph on p11 (sentence "We used the default settings for each…." ) to me suggests that cross-validation wasn't used to choose tuning parameters for the competitors, and that instead the default settings of the software were used!!! This would be really wacky and it may well be that I misunderstood what the authors meant -- but this issue needs to be corrected and/or clarified.