RELEASE HISTORY OF THE "sda" PACKAGE
========================================
CHANGES IN `sda' PACKAGE VERSION 1.3.3
- an import() statement has been added to NAMESPACE to address warnings
of R 3.1.0.
- added example scripts for Singh et al. (2001) and Khan et al. (2002)
gene expression data.
- now suggests "crossval" package for estimating prediction accuracy.
CHANGES IN `sda' PACKAGE VERSION 1.3.2
- sda() now also works with a single predictor (previously two predictors
were necessary).
- plot.sda.ranking() now has three new options to allow customization
(zeroaxis.col, ylab, and main).
CHANGES IN `sda' PACKAGE VERSION 1.3.1
- centroids() now also estimates class frequencies (in addition to
simply reporting the samples size per class). The frequencies are
estimated using a shrinkage approach (set lambda.freqs=0 for empirical
estimates). The pooled mean is now computed using the estimated
frequencies.
- catscore() now has a lambda.freqs argument and uses shrinkage
estimates of class frequencies to compute the scaling factor
(to use empirical scaling factors set lambda.freqs=0).
- the estimated frequencies returned by sda() are now contained
in the variable "freqs" (which previously was called "prior").
- in sda.ranking() there is now also a lambda.freqs argument
- in addition, sda.ranking() now offers three types of summary
statistics for ranking variables in the multi-class case.
CHANGES IN `sda' PACKAGE VERSION 1.3.0
- predict.sda() has been rewritten and is now *much* faster for large
numbers of test samples.
- the format of object returned by sda() has changed for more efficient
prediction. Note that it is *not* compatible with earlier versions.
CHANGES IN `sda' PACKAGE VERSION 1.2.4
- License file removed.
- Dependencies updated.
- plot.sda.ranking() is not based on "lattice" graphics any more
(new code contributed by Sebastian Gibb).
- sda() now allows to specify the shrinkage intensity for the
class frequencies.
CHANGES IN `sda' PACKAGE VERSION 1.2.3
- plot.sda.ranking() now checks for duplicated row names.
- feature.idx argument removed from predict.sda() function.
- sda.ranking() now allows to specify lambda and lambda.var
as in the catscore() function.
- sda() also has parameters to set lambda and lambda.var,
as well as shrink.freqs=TRUE/FALSE.
CHANGES IN `sda' PACKAGE VERSION 1.2.2
- centroids() function allows to specify the shrinkage intensity
for estimating the variances. Default is now shrinkage rather than
empirical estimates.
- catscore() also includes options to specifify shrinkage intensities.
The default is now using shrinkage rather empirical estimates.
- sda.ranking() now uses fdrtool to compute higher criticism scores
- in the output of sda(), the order of entries in the regularization
vector is now lambda, lambda.var, lambda.freqs.
CHANGES IN `sda' PACKAGE VERSION 1.2.1
- NAMESPACE file added
- updated requirements for "corpcor" and "entropy"
CHANGES IN `sda' PACKAGE VERSION 1.2.0
- requires now corpcor 1.6.0 and R version 2.10.0
- new function catscore()
- centroids() function has been streamlined and simplified
- updated documentation
- employs function crossprod.powcor.shrink() of corpcor
which leads to reduced memory imprint and increased speed
in functions catscore(), sda.ranking() and sda()
CHANGES IN `sda' PACKAGE VERSION 1.1.0
- new sda.ranking() function
- plot function for "sda.ranking" objects
- additional to FDR values computation of higher-criticism scores
- reference to Ahdesm\"aki and Strimmer (2009) paper added
- Singh et al. (2002) example data added
- improved help pages and examples
- the data khan.x is now on log-scale
CHANGES IN `sda' PACKAGE VERSION 1.0.3
- sda() now provides ranking of features.
- fdr values can optionally be computed for each feature.
- centroids() now reports number of samples and features.
- sda() function has been rewritten, and a bug introduced in
version 1.0.2 has been corrected.
CHANGES IN `sda' PACKAGE VERSION 1.0.2
- predict.sda() is now very much faster, and the object returned
by sda() needs much less memory.
- the centroids() function now additionally computes the pooled mean and
arbitrary powers of the correlation matrix (not just alpha=-1).
- the microarray data from Khan et al. 2001 are now used as example.
- bug fix: for shrinkage DDA the inverse correlation matrix is not
computed unnecessarily any more.
CHANGES IN `sda' PACKAGE VERSION 1.0.1
- new centroids() function to compute group-wise centroids,
(pooled variances), and inverse pooled correlations.
- uses the "collapse" option in corpcor >= 1.4.8 to save
memory when estimated correlation is diagonal (effectively
turning LDA into DDA if the estimated shrinkage intensity lambda=1).
CHANGES IN `sda' PACKAGE VERSION 1.0.0
This package implements LDA and DDA classification,
where the training of the classifier is done via Stein-type
shrinkage of frequencies, variances, and correlation.
This approach is particularly suitable for high-dimensional
classification.
This is the first public release (27 October 2008).