Changelog: ========== - EDAM and friend added - SVM Binaries no longer included (license) - Examples for calc.trans(), hmm.sop() updated - smrplot() no longer included (see below) - EDAM (incl. all related stuff and docs) optimized - fixes und improvements for shardsplot - partimat: argument plot.matrix = FALSE introduced to allow style like a scatterplot-matrix / pairs plot; some related changes to get docs / other arguments - triplot, tritrafo, quadplot, quadtrafo, centerlines added - shardsplot: + now utilizes the whole figure region + bugfix in expanded shardsplots + new args: xlab, ylab, xaxs, yaxs + corrected typos in docs - shardsplot: + Examples in EDAM() and shardsplot() fixed in resp. to legend() calls + box() added in shardsplot() - quadplot and friends: Minor doc fixes. - triplot and friends added - triplot doc fixes by KL, UL, CR, and thereafter *again* by UL. - shardsplot, countries: Zwei keine Fehler in den Manuals geändert - minor changes in: tritrafo + docs, predict.svmlight, quadplot docs, ucpm docs - quadplot "fixed" and several minor doc. changes ==== release: 0.2-1 (first one on CRAN) - quadplot arguments added and code improved to allow much more fine-tuning. - swapped order of returned arguments in predict.svmlight() - triperplines + docs added, triplot.R + docu (examples) changed accordingly - stepclass() can handle ucpm() measures now - stepclass() improvement fixed. improvement is now an absolute measure that works with arbitrary ucpm()s. priors have been defined for the null model. Defaults now to 0.05 rather than 0.95 Explicit argument "prior" has been removed, since the implementation war crude anyway. *Many* cosmetic changes - fix in ucpm(), failed if no colnames were given for the argument. - fix in svmlight / predict.svmlight: also works now if no path is given - svmlight() now returns values scaled via e.scal() - some cosmetic changes to svmlight and the help page - fixes for rda()'s optimization code - fixes for help files: NaiveBayes, b.scal, betascale - cosmetic changes to correpsonding code ==== release: 0.3-0 (second one on CRAN) - new argument legend.bg in drawparti - some more fixes for beta scaling - major fixes and extensions (regression, o-a-o) for svmlight - membercheck() function introduced (checks in ucpm and betascale) - NAMESPACE added ==== release: 0.3-2 - major bugfixes: *.matrix have been made NAMESPACE compatible stepclass, try-catching corrected predict.svmlight oao works for two classes ==== release: 0.3-3, 0.3-4 - minor fixes (e.g. to calm down CRAN maintainer) ==== release: 0.4-0 - first preparations for translations - some bug fixes, new args and better docs for plot functions - some regressions tests added - greedy.wilks and benchB3 added - nm (nearest mean) - meclight added - stepclass returns a formula - quadplot fixed now (setting "usr" appropriately), works now with together with fixed scatterplot3d_0.3-22 ==== release: 0.4-1 - stepclass has new argument min1var = TRUE (-> at least one variable in the model) - terms(....., data=data) ==== release: 0.4-2 - minor svmlight.Rd improvement - all.equal stuff for R-2.3.x changed ==== release: 0.5-0 - loclda and friends added ==== release: 0.5-1 - cat -> message - fixed links to Uschi's Diss - GPL 2 (but not newer!) - loclda speed improvement ==== release: 0.5-2 - Laplace correction for NaiveBayes ==== release: 0.5-3 - better computational properties in predict.NaiveBayes for many variables Thanks to Max Kuhn! ==== release: 0.5-4 - many fixes for the codetools checks ==== release: 0.6-0 - additions to shardsplot (for som display) by Dominik Reusser - new function level_shardsplot and helper rerange by Dominik Reusser - pvs, locpvs, corclust added ==== release: 0.6-1 - kmodes added ==== release: 0.6-2 - fix for numerics in predict.NaiveBayes by Karsten Luebke (bug reported by Feiming Chen) - K. Luebke's mail address changed ==== release: 0.6-3 - document several instances of "..." arguments that had empty argument descriptions - passing "..." to underlying print() functions for print.rda() ==== release: 0.6-4 - another fix for numerics in predict.NaiveBayes by Karsten Luebke - fix for wrong error message in case of 0 in 2nd column of tables for factors ==== release: 0.6-7 - fix for shardsplot color problems by Dominik Reusser ==== release: 0.6-8 - fix for data() without envir - fix for more than 100 factor levels in stepclass (reported by Ryan Brady) ==== release: 0.6-9 - new function woe() and its methods: Computes weight of evidence transform of factor variables for binary classification - fix: better error message: greedy.wilks does not support more than one grouping variable - fix: NaiveBayes work for explanatory variables that are factors with only 1 level (reported by Jeff Allard) - fix: break lines wider than 100 characters in Rd files ==== release: 0.6-10 - rename Changes-ToDo -> NEWS to make R CMD check happy - move combinat from Suggests to Imports - remove require calls for MASS ==== release: 0.6-11 - GermanCredit data copied vom evtree package which has been archived - depends on R >= 2.10.0 ==== release: 0.6-12 - new function cond.index for the "Calculation of Condition Indices for Linear Regression" - bugfix for NaiveBayes: predict would not work for one explan. variable ==== release: 0.6-13 - woe: speedup and extension for weighted observation (imports wtd.table from questionr) - kmodes: bugfix (wrong calculation for weights) speedup (additional fast = TRUE argument) [note: works on data frames instead of matrices (cf. ln. 3/83) and characters instead of integers (cf. ln. 42/43)] - corclust: extension to categorical variables (using Cramer's V); returned object now of class corclust; plot.corclust() added; additional functions xtractvars() and cvtree() [cutree() for variables] added that also returns within vs. between cluster correlations (not yet exported) - NaiveBayes update for new versions of R - predict.NaiveBayes gains support for logical explanatory variables thanks to Vincent R. Nijs - various fixes to make new checks happy - cvtree and xtractvars gain help files and are exported now. ==== release: 0.6-14 - make MKL check happy (print projection without orientation) - add tests for plots (we never diffed the ps files) ==== release: 0.6-15 - fix length(class) > 1 problem - fix Laplace correction thanks to Yan Li - add sanity check if not all required variables are provided - update reference output - adapt to changes in RNG (tests use the old RNG) ==== release: 1.7-0 - new version to indoicate this is pretty stable (for many years already) - add packages mentioned in Rd links to "Enhances" to avoid Notes - update test results for changes in the R graphics device - omit printing some lda test results to avoid eigenvector sign related diffs ==== release: 1.7-1 - fix some LaTeX isues (replace \& by & in Rd files) - fix issues when comparing classes to a length 1 vector - drawparti and hence partimat try to decide for a non random class in case of (almost) identical posteriors - predict.NaiveBayes now handles newdata correctly if factor structure does not match the training datas' - corclust: initialization of kor matrix fixed ==== release: 1.7-2 - rename internal function svmlight.file to svmlight_file to make S3 checks happy ==== release: 1.7-3 - fix lost braces in documentation - change tests not to produce any diffs To Do (at least in Uwe's opinion): ====== 1. disco must go into klaR and to CRAN! (e.g. for partimat) stepclass: - LATER: many rather long functions defined in stepclass() should be defined in the top level of the package (e.g. documenting them as "internal") - LATER: Rewrite it to be readable. - allow for *arbitrary* performance measures - allow for arbitrary predict functions rda: - In Main Procedure some loops within loops within ... with really bad memory allocatoion as in: new <- c(new, ......) e.g. lines 383 ff., the same with calls to cbind() - LATER: rda.default() - Many functions defined therein rather than in the package itself. Why? - LATER: another optimizer, e.g. optim()? quadplot: - rgl implementation triplot/quadplot: - check whether membership values are in [0,1]