Closed ramhiser closed 10 years ago
After much thought regarding what to do about alpha_k
in both the paper and in the code, I've opted to make it a parameter of convenience for flexible modeling.
Here's a brief blurb that will be added to the paper:
The addition of the parameter $\alpha_k$ is one of convenience and allows the flexibility of various covariance-matrix estimators proposed in the literature. In practice, we generally are not in estimating $\alpha_k$ as it would incorporate $K$ additional tuning parameters to estimate via cross-validation, which is counterproductive to our goal of improving the \emph{RDA} classifier's efficiency.
In the code, we will remove alpha_k
as an explicit numerical value passed as a parameter. Instead, we can incorporate some other argument, named shrinkage_family
or something similar. The possible values right now are:
ridge
: alpha_k = 1
to resemble ridge regressionconvex
: alpha_k = 1 - gamma
to resemble Friedman's estimator (need to define the range of gamma's carefully between 0 and 1, inclusively)More shrinkage_family
options could be added later if necessary, but for the HDRDA paper, this approach will greatly simplify things.
I suspect that one of the two shrinkage_family
arguments will outperform the other. How we determine this should be added to the paper, even if added as supplementary information. The convex
argument is possibly more straightforward because the range of candidate values of gamma
becomes [0, 1], whereas the values for the standard ridge
approach are unbounded. In the latter case, it's not always obvious which maximum value should be considered.
I added the shrinkage_type
argument to sparsediscrim::hdrda
. It has the two options given above. If convex
, the range of candidate values of gamma
is [0, 1].
If effective, demonstrate this and original approaches in paper?