[RsR] port of LIBRA toolbox to R

Mon Feb 5 18:49:14 CET 2007

(just a short comment, to show that we are still alive, although quite)

Thanks Tobias for your initiative and many, many thanks to Mia and Sabine.

As Martin already mentioned, you do not need to bother about the basic
MCD and LTS which are very efficiently implemented in R as native code
(FORTRAN) - see http://www.r-project.org/user-2006/Slides/Todorov.pdf
for a time comparison.

Also, I do not remember if Martin mentioned this, the cluster analysis
methods of LIBRA are included in the recommended R package 'cluster',
again in native C and FORTRAN code.

I have implementations of linear discriminant analysis (using not only
MCD, but also OGK M and S estimates, see the above presentation) as
well as of PCA, based on projection pursuit and MCD (ala ROBPCA) which
I intend to release soon in rrcov.

By the way, Martin when do you expect to deliver robustbase as a
recommended package? I still have not released a version of rrcov
without covMcd, ltsReg and related functions and requiring robustbase.

Best regards,
Valentin

On 2/4/07, Martin Maechler <maechler using stat.math.ethz.ch> wrote:
> Hi Tobias and everyone else
>    ((BTW: "anybody home?" , you've all been so quiet recently ..))
>
> >>>>> "TobiasV" == Tobias Verbeke <tobias.verbeke using gmail.com>
> >>>>>     on Thu, 01 Feb 2007 21:43:58 +0100 writes:
>
>     TobiasV> Hi Martin, Martin Maechler wrote:
>
>     >>>>>>> "Peter" == Peter Filzmoser
>     >>>>>>> <P.Filzmoser using tuwien.ac.at> on Mon, 22 Jan 2007
>     >>>>>>> 08:15:11 +0100 writes:
>     >>
>     Peter> Dear Tobias, This is very good news. Your work will
>     Peter> be an important contribution to the project.
>     >>  Indeed; thank you Tobias (and Mia and Sabine)!
>     >>
>     >> In the mean-time, I've studied a bit the "function list"
>     >> of LIBRA, and found that quite a few functions there seem
>     >> to be there just because matlab is not R ("base + stats +
>     >> MASS"); Several other ones have now been part of
>     >> 'robustbase' for a while.  Most of the remaining
>     >> functions where a port would seem very valuable are I
>     >> think related to robust PCA and versions thereof.
>
>     TobiasV> You're (of course) right that most of the basic
>     TobiasV> LIBRA functionality is already present in the named
>     TobiasV> R packages. As far as robust PCA is concerned, Mia
>     TobiasV> and Sabine also agree that this is the lion's part
>     TobiasV> of the job. They argued in the first place for the
>     TobiasV> inclusion of `robpca' [1] as it appears to stand
>     TobiasV> competition well amongst MCD-based methods.
>
>     TobiasV> Mia, Sabine and myself already had a discussion
>     TobiasV> function by function on what is and is not yet
>     TobiasV> available. Here is a summary which also reflects
>     TobiasV> priority (no logic in there) on my TODO list:
>
> Thanks a lot, that's very useful!
>
>     TobiasV> (1) For the moment I am working on the medcouple
>     TobiasV> and the skewness adjusted version of the boxplot
>     TobiasV> which would both be candidates for inclusion in
>     TobiasV> robustbase.
>
>     TobiasV> (2) Next, I plan to address `robpca', also
>     TobiasV> candidate for robustbase (cf. supra). Mia and
>     TobiasV> Sabine are reconsidering for the moment more
>     TobiasV> efficient algorithms for the cross-validated PRESS
>     TobiasV> value used that are not yet part of the current
>     TobiasV> version of MATLAB LIBRA. Together with robpca, some
>     TobiasV> diagnostic plots for robust PCA were developed
>     TobiasV> which could be added as well.
>
> Yes; though I think the computation and plotting should
> typically well separated, in typical S style
>
> Foo() computes an object of class "Foo" which has
> a print() [S3] or show() [S4] and a plot() method,
> and possibly a summary() {which produces a "summaryFoo" object that
> prints with a print.summary.Foo() methods}.
>
>     TobiasV> (3) Work towards integration (if not yet present)
>     TobiasV> in robustbase of multivariate linear models. A
>     TobiasV> design allowing the specification of different
>     TobiasV> robust covariance matrices to be used might offer
>     TobiasV> the flexibility R people are used to. This
>     TobiasV> corresponds to the (MCD-only) function `Mcdregres'
>     TobiasV> in LIBRA.
>
> good point and very welcome:  Traditional regression with lm()  works
> multi-variately "built-in".  From lmrob() this has been one of
> the several "TODO" items...
>
>     TobiasV> (4) robustified versions of traditional
>     TobiasV> classification methods (robust discriminant
>     TobiasV> analysis [`rda'] etc.). Should not be part of
>     TobiasV> robustbase and will be packaged separately as more
>     TobiasV> recent classification methods outperform these
>     TobiasV> traditional methods.
>
> Yes, indeed. I forgot to mention these in my first e-mail.
>
>     TobiasV> (5) robust versions of popular methods in
>     TobiasV> chemometrics such as `rsimpls', `rpcr', etc. These
>     TobiasV> again are no candidate for inclusion in robustbase
>     TobiasV> and are planned to be packaged separately.
>
>     TobiasV> Kind regards, Tobias
>
> I wish you a good start, and am glad to collaborate .. in about
> 5 weeks since I'm very busy just now before the DSC meeting in
> Auckland and subsequent long vacations...
>
> Martin
>
>
>     TobiasV> [1] Hubert, M., Rousseeuw, P.J., Vanden Branden,
>     TobiasV> K. (2005).  ROBPCA: a new approach to robust
>     TobiasV> principal component analysis, Technometrics, 47,
>     TobiasV> 64-79.
>
>     >> One question that may become interesting a bit later
>     >> (when speed & memory performance may be considered) is
>     >> how much of the code can be and should be built on the
>     >> fast-MCD code from robustbase.  I still have plans to
>     >> translate that to C and "export its API" since nowadays R
>     >> packages *can* export their C API, i.e. have other R
>     >> packages' C or R code directly call the C code of the
>     >> first package.
>     >>
>     >> Regards, Martin
>     >>
>     Peter> Tobias Verbeke wrote:
>     >> >> Dear R-SIG-Robust,
>     >> >>
>     >> >> I am happy to announce that Mia Hubert and Sabine >>
>     >> Verboven decided to let me port all of the LIBRA >>
>     >> functionality (cf.  >>
>     >> http://wis.kuleuven.be/stat/robust/LIBRA.html ) to R and
>     >> >> release the resulting work under GPL.
>     >> >>
>     >> >> I will discuss progress on the list not only to keep
>     >> you >> posted but also to make the best decisions
>     >> together in >> order to seamlessly integrate the
>     >> functions with the >> current framework.
>     >> >>
>     >> >> Everything should be available by October 2007.
>     >> >>
>     >> >> Best wishes, Tobias
>     >> >>
>     >>
>     >>
>     Peter> --
>     Peter> -------------------------------------------------------
>     Peter> From: Prof. Dr. Peter Filzmoser Dept. of Statistics &
>     Peter> Probability Theory Vienna University of Technology
>     Peter> Wiedner Hauptstrasse 8-10 A-1040 Vienna, Austria
>     Peter> Tel. +43 1 58801/10733 Fax. +43 1 58801/10799 E-mail:
>     Peter> P.Filzmoser using tuwien.ac.at Internet:
>     Peter> http://www.statistik.tuwien.ac.at/public/filz/
>     >>
>     Peter> _______________________________________________
>     Peter> R-SIG-Robust using r-project.org mailing list
>     Peter> https://stat.ethz.ch/mailman/listinfo/r-sig-robust
>     >>
>
>
>     TobiasV> --
>
>     TobiasV> Tobias Verbeke - Consultant Business & Decision
>     TobiasV> Benelux Rue de la révolution 8 1000 Brussels -
>     TobiasV> BELGIUM
>
>     TobiasV> +32 499 36 33 15
>     TobiasV> tobias.verbeke using businessdecision.com
>
> _______________________________________________
> R-SIG-Robust using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-robust
>