The R Project SVN R-packages

Things that were on my TODO list and have been accomplished
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ (newest 1st)

o Incorporate the PAM speedup patches by Matthias Studer (unige.ch).
  It's based on a good paper etc
  ---> ~/R/MM/Pkg-ex/cluster/pamonce-mail.txt
  ---> ~/R/MM/Pkg-ex/cluster/pamonce.R

o  daisy() -- now allows 'metric = "gower"'
      which allows to use "Gower's formula" also for the
      all-continuous case.
   The help page contains "Gower" much more visibly

o  Jan.2006:
   Create a new  silhoutte.clara(...., full = FALSE)
   where "full = TRUE" would compute all the silhouttes, not just those of
   the "best sample".

o  Dec.2005: need ./inst/CITATION  for getting a nice  citation()

o  May, 2004: New agnes(*, method = "flexible", par.method = *) to allow
   Lance-Williams formula.

o  Mar, 2004: When using a NAMESPACE, can rename `..dClass' in R/0aaa.R
   (and *.q): Renamed to 'dissiCl'

o  Dec, 2003: cutree(diana(), h = h0)
    gives complete non-sense without a warning
    -- the same as in S-plus ..



For R 1.7.0 :

  o agnes(), diana() and pam() have a new argument  `keep.diss'
    with a smart default.  If FALSE, the new behavior is  *NOT* to keep the
    dissimilarities !
     --> save space in result object
     --> somewhat changed summary/print output , i.e. minor NON-compatibility!

  o a) pam()$diss  is not a `proper' dissimilarity
       (class is only "dissimilarity", missing "dist") -> done Mo,17.3.03
    b) as.dist() is not generic and does too much for "dist" objects,
       and also for "dissimilarity" {<< these were fixed for R 1.7.0}
       and the completely wrong thing for the
       pam()$diss like "dissimilarity"s... (wrong $Size !)

  o plot(ylab / xlab) -- for agnes -> bannerplot() & pltree()
    ./R/plothier.q   ===> NOT DONE -- rather tell people {on the help()} to
    use the two underlying plot functions separately !

o Robert G:  " silhouette.default() is wrong "
  --> ok, I found this was on R-help, Feb.7 and fixed in my sources subsequently

Oct. 2002 {for "banner"; silhouette was earlier}:

 1) Silhouette und Banner Plot :
    Label vertical axis of horizontal barplot using HORIZONTAL text
    {i.e., par(las = 1) or axis(*, las = 1)}
    --> Should become an optional (but probably default)
    feature of barplot()!


July 2002:

  clara():  Must have allocation(?) bug in Fortran --> ../cluster_tests/clara.R

        --- bug (not allocation!) found and eliminated.

  o all [ agnes, clara, daisy, diana, fanny, pam ] now have something like

        valmisdat <- min(x2, na.rm=TRUE) - 0.5 #(double) VALue for MISsing DATa

    but this will go wrong as soon as  min(x2) < -5e15  !!!

   --> now using something more sensible

  o Generally:  clara(), pam(), agnes(), diana(),... should *not* keep the
      "diss" component (with all n^2/2 dissimilarities) in the result,
      by default when  n >= 50
      {i.e. get a new argument  `` keep.diss =  n < 50 '' }
        (done for 1.7.0, but default = n < 100)


  clara():  R/clara.q :  we transpose *large* x[,] and in Fortran
            src/clara.c  deal with x[] as vector anyway -- change this
        (done for 1.7.0)


May 2002:

  data/flower.R :  Big problem: V1..V3 should be binary, i.e. 0:1
  man/flower.Rd :  but as factors they become  1:2 !
        ==> the example (.. type = list(asymm = 1)
        has probably been wrong all the time -- compare to JSS  paper!
  --- no!  the changes to cluster-1.5.1 which test these,etc
      give no difference!

o clusplot() -- drawing MVE ellipses (-> .Fortran("spannel",..)) :
         S-plus has ellipses that just exactly contain all points.
         whereas our ellipses are *always* slightly too large.

o Provide a user function
     ellipsoidhull <- function(x, n = 201)
     = a combination of internal .Fortran("spannel",.), cov.wt() & ellipse()


o man/flower.Rd :  use \describe{  \item ... for the 8 variables

* OLDER

  2) help files for plot.*()
     --- new arguments are in *.Rd,
     but not really documented.

  5) a more thorough test of .Fortran("spannel", *) used in clusplot(*).

     -> done {new ellipsoidhull() }

-------------

./R/daisy.q  :

  Now that daisy objects inherit from "dist",  we can use  as.matrix() and
  other dist methods which can be *very* useful !
  --> also upgraded ./man/dissimilarity.object.Rd

  {remaining:  improve our code or the examples ?!}

 8) Make bannerplot() an (possibly later namespace internal ?)
    standalone function called from both  plot.agnes() and plot.diana(),
    even plot.mona()!