The main function kcca implements a general framework for k-centroids cluster analysis supporting arbitrary distance measures and centroid computation. Further cluster methods include hard competitive learning, neural gas, and QT clustering. There are numerous visualization methods for cluster results (neighborhood graphs, convex cluster hulls, barcharts of centroids, ...), and bootstrap methods for the analysis of cluster stability.
Changes in flexclust version 1.3-4
o Better documentation of differences of qtclust() to original algorithm.
o New function priceFeature() for artificial market segments.
o .rda files for data now use xz compression.
o New data sets auto and volunteers.
o Shaded boxplots and barcharts.
Changes in flexclust version 1.3-3
o bootFlexclust() did not work for k of length 1.
o Use package parallel instead of snow and multicore.
o Argument FUN of stepFlexclust() can be a character string naming the function.
o kccaFamily("angle") supports weights.
o New conversion method as.kcca.skmeans().
o Use cor() instead of .Internal() cor.
Changes in flexclust version 1.3-2
o It was not documented that argument k in kcca() and cclust() can also be a vector of initial cluster assignments (in addition to number of clusters and matrix of centroids). The feature has been available for a long time.
o The line segments to the poulation mean in the barchart() method can now have a diffent color than the corresponding points.
o It is now documented and checked that parallel processing with snow currently works only for bootFlexclust(), not stepFlexclust().
o New initialization strategy kmeans++ for kcca() and cclust(), invoked using the control argument. Users can also specify and implement their own initialization strategies.
Changes in flexclust version 1.3-1
o Bugfix in parallel processing with snow.
o Do not run example with snow by default (does not work on Solaris), snow test is now private.
Changes in flexclust version 1.3-0
o Added support for parallel processing with package snow in addition to multicore.
o Added data from election for German parliament 2009.
o barchart(clobj, mcol=NULL) did not work.
o New conversion method from kccasimple to S3 kmeans.
Changes in flexclust version 1.2-2
o New randIndex methods for integer arguments (cluster membership vectors).
o New function as.kcca.hclust().
o Fixed two bugs in the CITATION file.
Changes in flexclust version 1.2-1
o New arguments seed and multicore for bootFlexclust(), which now by default uses multicore for parallel processing if available. Note that this changes handling of random numbers (we need parallel streams), so results with the same seed are not the same as for previous versions of the package.
o New wrapper stepcclust() for stepFlexclust(..., FUN=cclust).
Changes in flexclust version 1.2-0
o New functions propBarchart(), shadow(), Silhouette(), shadowStars() and stripes(), see corresponding help files for details.
o List subsetting operator [] can now be used as a shortcut for getModel() on stepFlexclust objects.
o New argument legend in barchart() method. Percentages in panel strips now rounded to next integer.
o New arguments seed and multicore for stepFlexclust(), which now by default uses multicore for parallel processing if available. Note that this changes handling of random numbers (we need parallel streams), so results with the same seed are not the same as for previous versions of the package.
Changes in flexclust version 1.1-2
o Set "ZipData: no" in DESCRIPTION.
Changes in flexclust version 1.1-1
o Bugfix in barchart().
Changes in flexclust version 1.1-0
o New generic randIndex() for computation of the Rand Index for partition agreement.
o New function bootFlexclust() to bootstrap cluster stability.
o show() method for kcca objects reports NAs in cluster assignment only if there are any.
Changes in flexclust version 1.0-0
o New data set with German Bundestag election results.
Changes in flexclust version 0.99-1
o kcca() now chooses the random set of initial centroids only from the complete observations in x, this allows the usage of distance and centroid computation functions which can deal with missing values.
o Bugfix: cclust segfaulted for manhattan distance when clusters got empty.
o Don't use the show() method to plot projected Axes, it only confuses people (including myself).
o qtclust() now has an argument
kcca' instead ofsimple' with
slightly different meaning (at the end, run kcca() to convergence
rather than only one iteration). This also fixes a bug which
reported differnt cluster sizes in the print() and summary()
method for the return objects of qtclust().
o The never really used support for similarity measures has been dropped to make the code cleaner and easier to maintain.
Changes in flexclust version 0.99-0
o qtclust() searched only for clusters with size>min.size rather than the documented size>=min.size.
o qtclust() now by default returns objects of class "kccasimple" which contains only a subset of the slots of regular "kcca" objects (and hence are much faster to compute). The new default was chosen because if the radius is chosen too small, the number of clusters can easily be very large.
o show() now prints only summary statistics of cluster sizes if there are more than 20 clusters.
o The 'project' argument of the plot() method now also handles 'lda' objects correctly.
o The barchart() method has several new arguments and is now much more flexible.
o Fixed two typos in data set dentitio ("ferret" was "ferrer", "pygmy bat" was "pigmy bat").
o Improved handling of color of cluster centroid labels in the plot method.
o New "medium" color palette.
o The was a bug in stepFlexclust() when using families with a data pre-processing function like kccaFamily("angle").
o The data matrix x can now be stored in the return object and is then used automatically by the plot() method.
o New generic randomTour() with methods for data matrices and flexclust objects.
o Renamed cluster() to clusters() to avoid conflict with cluster() from package survival
o The default hulls for clusters are now convex hulls rather than ellipses, which can be misleading for non-Gaussian data.
Changes in flexclust version 0.9-1
o new argument newdata for cluster()
Changes in flexclust version 0.9-0
o kcca() and cclust() can now optionally return objects of class "kccasimple" which contains only a subset of the slots of regular "kcca" objects (basically only centers and cluster membership). Computation is faster, but no plotting methods are available.
o Cluster membership vectors now preserve the rownames of the original data.
o Fixed a bug in Jaccard distance computation and random break of ties for cluster assignment, both reported by Christian Buchta.
o New barchart() method for visualization of centroids.
o Utility function flxColors() can be used to access the flexclust color palette.
Changes in flexclust version 0.8-1
o Fixed a bug in the definition of class "flexclustControl".
Changes in flexclust version 0.8-0
o List ellipse as suggested package, use only when available.
o Class "flexclust" has a new slot "clusinfo", which is modeled after the same slot of partition objects in package cluster. This replaces slots "size" and "withindist".
o New generic info() can be used to obtain various pieces of information about a clustering.
o New summary() method lists contents of clusinfo slot, show() method only cluster sizes.
o Added details of neural gas learning paramters to flexclustControl-class.Rd and fixed a type (number of parameters is 4, not 3).
o New methods cluster() and parameters().
o New function projAxes() to add arrows for original coordinate axes to a projection plot.
o New pairs() method for matrix of neighborhood graphs.
o New function dist2().
o Improved performance of qtclust() and distCor().
o barplot() now plots clusters by default in the same colors as plot().
o Correct citation of the package is the forthcoming CSDA paper.
First version released on CRAN: 0.7-0