Gaussian finite mixture models fitted via EM algorithm for model-based clustering, classification, and density estimation, including Bayesian regularization, dimension reduction for visualisation, and resampling-based inference.
Version 5.4.1 (NOT ON CRAN) o Add parametric bootstrap option (type = "pb") in MclustBootstrap(). o Add the options to get averages of resampling distributions in summary.MclustBootstrap() and to plot resampling-based confidence intervals in plot.MclustBootstrap(). o Add function catwrap() for wrapping printed lines at getOption("width") when using cat(). o mclust.options() now modify the variable .mclust in the namespace of the package, so it should work even inside an mclust function call. o Fix a bug in covw() when normalize = TRUE. o Fix a bug in estepVEV() estepVEE() when parameters contains Vinv. o Fix a bug in plotDensityMclustd() when drawing marginal axes. o Fix a bug in summary.MclustDA() when computing classification error in the extreme case of a minor class of assignment. o Fix a bug in the initialisation of mclustBIC() when a noise component is present for 1-dimensional data. o fix bugs in some examples documenting clustCombi() and related functions.
Version 5.4 o Model-based hierarchical clustering used to start the EM-algorithm is now based on the scaled SVD transformation proposed by Scrucca and Raftery (2016). This change is not backward compatible. However, previous results can be easily obtained by issuing the command: mclust.options(hcUse = "VARS") For more details see help("mclust.options"). o Added 'subset' parameter in mclust.options() to control the maximal sample size to be used in the initial model-based hierarchical phase. o predict.densityMclust() can optionally returns the density on a logarithm scale. o Removed normalization of mixing proportions for new models in single mstep. o Internal rewrite of code used by packageStartupMessage(). o Fix a small bug in MclustBootstrap() in the univariate data case. o Fix bugs when both the noise and subset are provided for initialization. o Vignette updated to include references, startup message, css style, etc. o Various bugs fix in plotting methods when noise is present. o Update references in citation() and man pages.
Version 5.3 (2017-05) o added gmmhd() function and relative methods. o added MclustDRsubsel() function and relative methods. o added option to use subset in the hierarchical initialization step when a noise component is present. o plot.clustCombi() presents a menu in interactive sessions, no more need of data for classification plots but extract the data from the 'clustCombi' object. o added combiTree() plot for 'clustCombi' objects. o clPairs() now produces a single scatterplot in the bivariate case. o fix a bug in imputeData() when seed is provided. Now if a seed is provided the data matrix is reproducible. o in imputeData() and imputePairs() some name of arguments have been modified to be coherent with the rest of the package. o added functions matchCluster() and majorityVote(). o rewriting of print and summary methods for 'clustCombi' class object. o added clustCombiOptim(). o fix a bug in randomPairs() when nrow of input data is odd. o fix a bug in plotDensityMclust2(), plotDensityMclustd() and surfacePlot() when a noise component is present.
Version 5.2.3 (2017-03) o added native routine registration for Fortran code. o fix lowercase argument PACKAGE in .Fortran() calls.
Version 5.2.2 (2017-01) o fix a bug in rare case when performing an extra M step at the end of EM algorithm.
Version 5.2.1 (2017-01) o replaced "structure(NULL, *)" with "structure(list(), *)"
Version 5.2 (2016-03) o added argument 'x' to Mclust() to use BIC values from previous computations to avoid recomputing for the same models. The same argument and functionality was already available in mclustBIC(). o added argument 'x' to mclustICL() to use ICL values from previous computations to avoid recomputing for the same models. o corrected a bug on plot.MclustBootstrap for the "mean" and "var" in the univariate case. o modified uncertainty plots. o introduction of as.Mclust and as.densityMclust to convert object to specific mclust classes. o solved a numerical accuracy problem in qclass when the scale of x is (very) large by making the tolerance eps scale dependent. o use transpose subroutine instead of non-Fortran 77 TRANSPOSE function in mclustaddson.f o predict.Mclust and predict.MclustDR implement a more efficient and accurate algorithm for computing the densities.
Version 5.1 (2015-10) o fix slow convergence for VVE and EVE models. o fix a bug in orientation for model VEE. o add an extra M-step and parameters update in Mclust call via summaryMclustBIC.
Version 5.0.2 (2015-07) o add option to MclustBootstrap for using weighted likelihood bootstrap. o add a plot method to MclustBootstrap. o add errorBars function. o add clPairsLegend function. o add covw function. o fix rescaling of mixing probabilities in new models. o bug fixes.
Version 5.0.1 (2015-04) o bug fixes. o add print method to hc.
Version 5.0.0 (2015-03) o added the four missing models (EVV, VEE, EVE, VVE) to the mclust family. A noise component is allowed, but no prior is available. o added mclustBootstrapLRT function (and print and plot methods) for selecting the number of mixture components based on the bootstrap sequential likelihood ratio test. o added MclustBootstrap function (and print and summary methods) for performing bootstrap inference. This provides standard errors for parameters and confidence intervals. o a "A quick tour of mclust" vignette is included as html generated using rmarkdown and knitr. Older vignettes are included as other documentation for the package. o modified arguments to mvn2plot to control colour, lty, lwd, and pch of ellipses and mean point. o added functions emX, emXII, emXXI, emXXX, cdensX, cdensXII, cdensXXI, and cdensXXX, to deal with single-component cases, so calling the em function works even if G = 1. o small changes to icl.R, now icl is a generic method, with specialized methods for 'Mclust' and 'MclustDA' objects. o bug fixes for transformations in the initialization step when some variables are constant (i.e. the variance is zero) or a one-dimensional data is provided. o change the order of arguments in hc (and all the functions calling it). o small modification to CITATION file upon request of CRAN maintainers. o small bug fixes.
Version 4.4 (2014-09) o add option for using transformation of variables in the hierarchical initialization step. o add quantileMclust for computing the quantiles from a univariate Gaussian mixture distribution. o bug fixes on summaryMclustBIC, summaryMclustBICn, Mclust to return a matrix of 1s on a single column for z even in the case of G = 1. This is to avoid error on some plots. o pdf files (previously included as vignettes) moved to inst/doc with corresponding index.html.
Version 4.3 (2014-03)
o bug fix for logLik.MclustDA() in the univariate case.
o add argument "what" to predict.densityMclust() function for choosing
what to retrieve, the mixture density or component density.
o hc function has an additional parameter to control if the original
variables or a transformation of them should be used for
o included "hcUse" in mclust.options to be passed as default to hc().
o original data (and class for classification models) are stored in the
object returned by the main functions.
o add component "hypvol"" to Mclust object which provide the hypervolume
of the noise component when required, otherwise is set to NA.
o add a warning when prior is used and BIC returns NAs.
o bug fixes for summary.Mclust(), print.summary.Mclust(), plot.Mclust()
and icl() in the case of presence of a noise component.
o some plots on plot.MclustDR() require plot.new() before calling
o bug fixes for MclustDR() when p=1. o correction to Mclust man page. o bug fixes.
Version 4.2 (2013-07) o fix bug in sim* functions when no obs are assigned to a component. o MclustDA allows to fit a single class model. o fix bug in summary.Mclust when a subset is used for initialization. o fix a bug in the function qclass when ties are present in quantiles, so it always return the required number of classes. o various small bug fixes.
Version 4.1 (2013-04) o new icl function for computing the integrated complete-data likelihood o new mclustICL function with associated print and plot methods o print.mclustBIC shows also the top models based on BIC o modified summary.Mclust to return also the icl o rewrite of adjustedRandIndex function. This version is more efficient for large vectors o updated help for adjustedRandIndex o modifications to MclustDR and its summary method o changed behavior of plot.MclustDR(..., what = "contour") o improved plot of uncertainty for plot.MclustDR(..., what = "boundaries") o corrected a bug for malformed GvHD data o corrected version of qclass for selecting initial values in case of 1D data when successive quantiles coincide o corrected version of plot BIC values when only a single G component models are fitted o various bug fixes
Version 4.0 (2012-08) o new summary and print methods for Mclust. o new summary and print methods for densityMclust. o included MclustDA function and methods. o included MclustDR function and methods. o included me.weighted function. o restored hierarchical clustering capability for the EEE model (hcEEE). o included vignettes for mclust version 4 from Technical Report No. 597 and for using weights in mclust. o adoption of GPL (>= 2) license.
Version 3.5 (2012-07) o added summary.Mclust o new functions for plotting and summarizing density estimation o various bug fixes o clustCombi (code and doc provided by Jean-Patrick Baudry) o bug fix: variable names lost when G = 1
Version 3.4.11 (2012-01) o added NAMESPACE
Version 3.4.10 (2011-05) o removed intrinsic gamma
Version 3.4.9 (2011-05) o fixed hypvol function to avoid overflow o fixed hypvol helpfile value description o removed unused variables and tabs from source code o switched to intrinsic gamma in source code o fixed default warning in estepVEV and mstepVEV
Version 3.4.8 (2010-12) o fixed output when G = 1 (it had NA for the missing "z" component)
Version 3.4.7 (2010-10) o removed hierarchical clustering capability for the EEE model (hcEEE) o The R 2.12.0 build failed due to a 32-bit Windows compiler error, forcing removal of the underlying Fortran code for hcEEE from the package, which does not contain errors and compiles on other platforms.
Version 3.4.6 (2010-08) o added description of parameters output component to Mclust and o summary.mclustBIC help files
Version 3.4.5 (2010-07) o added densityMclust function
Version 3.4.4 (2010-04) o fixed bug in covariance matrix output for EEV and VEV models
Version 3.4.3 (2010-02) o bug fixes
Version 3.4.2 (2010-02) o moved CITATION to inst and used standard format o BibTex entries are in inst/cite o fixed bug in handling missing classes in mclustBIC o clarified license wording
Version 3.4.1 (2010-01) o corrected output description in mclustModel help file o updated mclust manual reference to show revision
Version 3.4 (2009-12) o updated defaultPrior help file o added utility functions for imputing missing data with the mix package o changed default max # of mixture components in each class from 9 to 3
Version 3.3.2 (2009-10) o fixed problems with \cr in mclustOptions help file
Version 3.3.1 (2009-06) o fixed plot.mclustBIC/plot.Mclust to handle modelNames o changed "orientation" for VEV, VVV models to be consistent with R eigen() and the literature o fixed some problems including doc for the noise option o updated the unmap function to optionally include missing groups
Version 3.3 (2009-06) o fixed bug in the "errors" option for randProj o fixed boundary cases for the "noise" option
Version 3.2 (2009-04) o added permission for CRAN distribution to LICENSE o fixed problems with help files found by new parser o changed PKG_LIBS order in src/Makevars o fixed Mclust to handle sampling in data expression in call
Version 3.1.10 (2008-11) o added EXPR = to all switch functions that didn't already have it
Version 3.1.9 (2008-10) o added pro component to parameters in dens help file o fixed some problems with the noise option
Version 3.1.1 (2007-03) o Default seed changed in sim functions. o Model name check added to various functions. o Otherwise backward compatible with version 3.0
Version 3.1 (2007-01) o Most plotting functions changed to use color. o Mclust/mclustBIC fixed to work with G=1 o Otherwise backward compatible with version 3.0.
Version 3.0 (2006-10) o New functionality added, including conjugate priors for Bayesian regularization. o Backward compatibility is not guaranteed since the implementation of some functions has changed to make them easier to use or maintain.