News

Infinite Mixtures of Infinite Factor Analysers

IMIFA v2.1.0 - (7^th release [minor update]: 2019-02-04)

New Features

• mgpControl gains the arguments cluster.shrink and sigma.hyper:
  • cluster.shrink governs invocation of cluster shrinkage MGP hyperparameter for MIFA/OMIFA/IMIFA methods.
  • sigma.hyper controls the gamma hyperprior on this parameter. The posterior mean is reported, where applicable.
  • Full conditionals for loadings and local/column shrinkage MGP hyperparameters edited accordingly.
• Allowed the Dirichlet concentration parameter alpha to be learned via MH steps for the OM(I)FA models.
  • Also allowed diminishing adaptation to tune the log-normal proposal to achieve a target acceptance rate.
  • Thus bnpControl args. learn.alpha, alpha.hyper, zeta, & tune.zeta become relevant for OM(I)FA models.
• New posterior predictive model checking approach added to get_IMIFA_results (with associated plots):
  Posterior Predictive Reconstruction Error (PPRE) compares bin counts of the original data with corresponding
  counts for replicate draws from the posterior predictive distribution using a standardised Frobenius norm.
• Added new function scores_MAP to decompose factor scores summaries
  from get_IMIFA_resuls into submatrices corresponding to the MAP partition.
• Added new wrapper function sim_IMIFA_model to call sim_IMIFA_data using
  the estimated parameters from fitted Results_IMIFA objects.
• New get_IMIFA_results arg. vari.rot allows loadings templates to be varimax rotated,
  prior to Procrustes rotation, for more interpretable solutions (defaults to FALSE).
• New plot.Results_IMIFA argument common governing plot.meth="means" plots (details in documentation).

Improvements

• New hyperparameter/argument defaults:
  • sigma.mu defaults to 1 s.t. the hypercovariance is the identity for the prior on the means;
    old behaviour (using the diagonal of the sample covariance matrix) recoverable by specifying sigma.mu=NULL.
  • prec.mu defaults to 0.01 s.t. the prior on the cluster means is flat by default.
  • learn.d defaults to TRUE s.t. a PYP prior is assumed for IM(I)FA models by default.
  • alpha.hyper now has a larger hyper-rate by default, to better encourage clustering.
  • alpha.d1 & alpha.d2 now set to 2.1/3.1 rather than 2/6 to discourage exponentially fast shrinkage.
  • z.init now defaults to "hc": model-based agglomerative hierarchical clustering.
• Overhauled psi_hyper (details in documentation) for:
  • N <= P data where the sample covariance matrix is not invertible.
  • type="isotropic" uniquenesses.
• Args. scores & loadings can now be supplied to sim_IMIFA_data directly;
  new arg. non.zero controls the # effective factors (per column & cluster) when loadings are instead simulated.
• Sped-up 2^nd label-switching move for IM(I)FA models (accounting for empty clusters).
• Args. for hc can now be passed when init.z="mclust" also
  (previously only "hc"), thus controlling how Mclust is itself initialised.
• Allowed criterion to be passed via ... in mixfaControl to choose between
  mclustBIC/mclustICL to determine optimum model to initialise with when
  z.init="mclust" & also sped-up mclust initialisation in the process.
• Added stop.AGS arg. to mgpControl: renamed adapt.at to start.AGS for consistency.
• Added start.zeta & stop.zeta options to tune.zeta argument in bnpControl.
• Allowed user-supplied breaks in the plotting functions mat2cols & heat_legend.
• Initial cluster sizes are now shown in order to alert users to potentially bad starting values.
• Added utility function pareto_scale().

Bug Fixes

• Fixed factor scores & error metrics issues in get_IMIFA_results for clustering methods:
  • Fixed storage of scores for infinite factor methods - now corresponds to samples where the
    largest cluster-specific number of factors is >= the max of the modal estimates of the same
    (previously samples where any cluster has >= the corresponding modal estimate were used):
    thus, valid samples for computing error metrics also fixed and Procrustes rotation also sped-up.
  • Other Procrustes rotation fixes to account for label-switching.
  • Other Procrustes rotation fixes specific to the IMFA/OMFA methods.
• range.G and trunc.G defaults fixed, especially for small sample size settings.
• Slight label-switching fixes when zlabels are supplied to get_IMIFA_results;
  posterior confusion matrix, cluster sizes vector, and the sampled labels themselves effected.
• Prevented unnecessary Procrustes rotation for single-factor components, thus fixing some bugs.
• Fixed initialisation of uniquenesses to account for all four settings of uni.type.
• Allowed conditioning on iterations with all components populated for M(I)FA models in get_IMIFA_results.
• Accounted for 1-component IM(I)FA/OM(I)FA models in get_IMIFA_results.
• Fixed handling of empty components when simulating cluster labels from priors in mcmc_IMIFA & sim_IMIFA_data.
• Ensured no. of factors Q cannot exceed no. of observations in the corresponding cluster in sim_IMIFA_data.
• Slight speed-up to updating MGP hyperparameters in the presence of empty MIFA/OMIFA/IMIFA components.
• Slight speed-up to sampling cluster labels with slice indicators for IM(I)FA models.
• Explicitly allowed Pitman-Yor special case where alpha=0 for IM(I)FA models;
  added related controls on spike-and-slab prior for discount when fixing alpha<=0.
• Allowed full range of hc model types for initialisation purposes via ... in mixfaControl.
• Clarified dimnames of get_IMIFA_results output in x$Loadings & x$Scores.
• Fixed storage switches & iteration indices to better account for burnin=0.
• Fixed plotting of exact zeros in posterior confusion matrix.
• Fixed plotting posterior mean loadings heatmap when one or more clusters have zero factors.
• Fixed plotting scores for (I)FA models due to bug in previous update, esp. with zlabels supplied.
• Fixed show_IMIFA_digit to better account for missing pixels &/or the data having been centered/scaled.
• Fixed simulation of psi when not supplied to sim_IMIFA_data to IG rather than GA.
• Fixed bug preventing Q to be supplied to get_IMIFA_results for infinite factor methods.
• Fixed y-axis labelling in uncertainty type plots when plot.meth="zlabels".
• Small fixes to function show_digit.
• Better handling of tied model-selection criteria in get_IMIFA_results.
• Procrustes now works when X has fewer columns than Xstar.
• Minor cosmetic change for overplotting scores & loadings in trace & density plots.
• Edited Ledermann and related warnings to account for case of isotropic uniquenesses.
• Tidied indentation/line-breaks for cat/message/warning calls for printing clarity.
• Corrected IMIFA-package help file (formerly just IMIFA).
• Edited CITATION file and authorship.

IMIFA v2.0.0 - (6^th release [major update]: 2018-05-01)

Major Changes

• Simplified mcmc_IMIFA by consolidating arguments using new helper functions (with defaults):
  • Args. common to all factor-analytic mixture methods & MCMC settings supplied via mixfaControl.
  • MGP & AGS args. supplied via mgpControl for infinite factor models.
  • Pitman-Yor/Dirichlet Process args. supplied via bnpControl for infinite mixture models.
  • Storage switch args. supplied via storeControl.
  • New functions also inherit the old documentation for their arguments.

New Features

• Posterior predictive checking overhauled: now MSE, RMSE etc. between empirical & estimated covariance
  matrices are computed for every retained iteration so uncertainty in these estimates can be quantified:
  • Can be switched on/off via the error.metrics argument to get_IMIFA_results.
  • Can be visualised by supplying plot.meth="errors" to plot.Results_IMIFA.
  • For methods which achieve clustering, the 'overall' covariance matrix
    is now properly computed from the cluster-specific covariance matrices.
  • Same metrics also evaluated at posterior mean parameter estimates & for final sample where possible.
• mixfaControl gains the arg. prec.mu to control the degree of flatness of the prior for the means.
• Posterior confusion matrix now returned (get_IMIFA_results) & visualisable (plot.Results_IMIFA,
  when plot.meth="zlabels"), via new function post_conf_mat, to further assess clustering uncertainty.
• Added new type of clustering uncertainty profile plot in plot.Results_IMIFA when plot.meth="zlabels".
• For convenience, get_IMIFA_results now also returns the last valid samples for parameters of interest,
  after conditioning on the modal G & Q values and accounting for label switching and Procrustes rotation.
• plot.Results_IMIFA gains new arg. show.last that replaces any instance of showing the posterior mean
  with the last valid sample instead (i.e. when plot.meth="means" or plot.meth="parallel.coords").
• Added ability to constrain mixing proportions across clusters using equal.pro argument for M(I)FA models:
  Modified PGMM_dfree accordingly and forced non-storage of mixing proportions when equal.pro is TRUE.
• All methods now work for univariate data also (with apt. edits to plots & uniqueness defaults etc.).
  sim_IMIFA_data also extended to work for univariate data, as well as sped-up.

Improvements

• Retired args. nu & nuplus1 to mgpControl, replaced by ability to specify more general gamma prior,
  via new phi.hyper arg. specifying shape and rate - mgp_check has also been modified accordingly.
• Zsimilarity sped-up via the comp.psm & cltoSim functions s.t. when # observations < 1000.
• Matrix of posterior cluster membership probabilities now returned by get_IMIFA_results.
• Modified AGS to better account for when the number of group-specific latent factors shrinks to zero.
• psi.alpha no longer needs to be strictly greater than 1, unless the default psi.beta is invoked;
  thus flatter inverse gamma priors can now be specified for the uniquenesses via mixfaControl.
• Added "hc" option to z.init to initialise allocations via hierarchical clustering (using mclust::hc).
• Allowed optional args. for functions used to initialise allocations via ... in mixfaControl.
• Added mu argument to sim_IMIFA_data to allow supplying true mean parameter values directly.
• Standard deviation of aicm/bicm model selection criteria now computed and returned.
• Speed-ups due to new Rfast utility functions: colTabulate & matrnorm.
• Speed-ups due to utility functions from matrixStats, on which IMIFA already depends.
• Slight improvements when adapt=FALSE for infinite factor models with fixed high truncation level.
• Misclassified observations now highlighted in 1^st type of uncertainty plot in Plot.Results_IMIFA,
  when plot.meth="zlabels" and the true zlabels are supplied.
• mixfaControl gains arg. drop0sd to control removal of zero-variance features (defaults to TRUE).
• heat_legend gains cex.lab argument to control magnification of legend text.
• mat2cols gains the transparency argument.
• Edited PGMM_dfree to include the 4 extra models from the EPGMM family.

Bug Fixes

• Supplying zlabels to get_IMIFA_results will now match the cluster labels and parameters to
  the true labels even if there is a mismatch between the number of clusters in both.
• Similarly, supplying zlabels to plot.Results_IMIFA when plot.meth="zlabels" no longer does
  any matching when printing performance metrics to the screen - previously this caused confusion
  as associated parameters were not also permuted as they are within get_IMIFA_results: now
  plot(get_IMIFA_results(sim), plot.meth="zlabels", zlabels=z) gives different results from
  plot(get_IMIFA_results(sim, zlabels=z), plot.meth="zlabels") as only the latter will permute.
• Accounted for errors in covariance matrix when deriving default sigma.mu & psi.beta values.
• Accounted for missing empirical covariance entries within get_IMIFA_results.
• Fixed model selection in get_IMIFA_results for IMFA/OMFA models when range.Q is a range.
• Fixed calculation of aicm, bicm and dic criteria: all results remain the same.
• Fixed support of Ga(a, b) prior on alpha when discount is being learned.
• Fixed bug preventing uni.prior="isotropic" when uni.type is (un)constrained.
• Fixed treatment of exact zeros when plotting average clustering similarity matrix.
• Fixed tiny bug when neither centering nor scaling (of any kind) are applied to data within mcmc_IMIFA.
• Fixed plotting of posterior mean scores when one or more clusters are empty.
• Fixed bug with default plotting palette for data sets with >1024 variables.
• Fixed bug with label switching permutations in get_IMIFA_results when there are empty clusters.
• Fixed print and summary functions for objects of class IMIFA and Results_IMIFA.
• Fixed calculating posterior mean zeta when adaptively targeting alpha's optimal MH acceptance rate.
• Allowed alpha be tiny for (O)M(I)FA models (provided z.init != "priors" for overfitted models).
• Normalised mixing proportions in get_IMIFA_results when conditioning on G for IM(I)FA/OM(I)FA models.
• New controls/warnings for excessively small Gamma hyperparemeters for uniqueness/local shrinkage priors.
• Clarified recommendation in MGP_check that alpha.d2 be moderately large relative to alpha.d1.
• Ensured sigma.mu hyperparameter arg. is always coerced to diagonal entries of a covariance matrix.
• Transparency default in plot.Results_IMIFA now depends on device's support of semi-transparency.
• Replaced certain instances of is.list(x) with inherits(x, "list") for stricter checking.
• Added check.margin=FALSE to calls to sweep.
• Ledermann, MGP_check, and PGMM_dfree are now properly vectorised.

Miscellaneous Edits

• Added USPSdigits data set (training and test),
  with associated utility functions show_digit and show_IMIFA_digit.
• Optimised compression of olive, coffee and vignette data and used LazyData: true.
• Added call.=FALSE to stop() messages and immediate.=TRUE to certain warning() calls.
• Removed dependency onadrop, e1071, graphics, grDevices, plotrix, stats & utils libraries.
• Reduced dependency on Rfast w/ own versions of colVars, rowVars, & standardise.
• Added utility function IMIFA_news for accessing this NEWS file.
• Added CITATION file.
• Extensively improved package documentation:
  • Added Collate: field to DESCRIPTION file.
  • Added line-breaks to usage sections of multi-argument functions.
  • Consolidated help files for G_expected & G_variance.

IMIFA v1.3.1 - (5^th release [patch update]: 2017-07-07)

Bug Fixes

• Fixed bug preventing M(I)FA models from being treated as (I)FA models when range.G contains 1.
• Fixed bug preventing get_IMIFA_results from working properly when true labels are NOT supplied.

IMIFA v1.3.0 - (4^th release [minor update]: 2017-06-22)

New Features

• Added options "constrained" & "single" to mcmc_IMIFA's uni.type argument:
  as well as being either diagonal or isotropic (UUU / UUC), uniquenesses can now further be
  constrained across clusters (UCU / UCC), with appropriate warnings, defaults, checks,
  initialisations, computation of model choice penalties, and plotting behaviour in all 4 cases.
• mcmc_IMIFA gains the tune.zeta argument, a list of heat, lambda & target parameters, to invoke
  diminishing adaptation for tuning the uniform proposal to achieve a target acceptance rate when alpha
  is learned via Metropolis-Hastings when the Pitman-Yor Process prior is employed for the IM(I)FA models.

Improvements

• (I)FA models sped up by considering uniquenesses under 1-cluster models as "constrained" or "single",
  rather than previously "unconstrained" or "isotropic", utilising pre-computation and empty assignment.
• Previously hidden functions improved, exported and documented with examples:
  is.cols, Ledermann, Procrustes & shift_GA.
• is.posi_def gains make argument, merging it with previously hidden function .make_posdef:
  Thus the 'nearest' positive-(semi)definite matrix and the usual check can be returned in a single call.
• Sped-up sampling IM(I)FA labels, esp. when 'active' G falls to 1, or the dependent slice-sampler is used:
  log.like arg. removed from gumbel_max; function stands alone, now only stored log-likelihoods computed.
• psi argument added to sim_IMIFA_data to allow supplying true uniqueness parameter values directly.

Bug Fixes

• Used bw="SJ" everywhere density is invoked for plotting (bw="nrd0" is invoked if this fails).
• Fixed initialisation of uniquenesses for isotropic (I)FA models.
• Fixed parallel coordinates plot axes and labels for all isotropic uniquenesses plots.
• Fixed adaptation for MIFA/OMIFA/IMIFA models when all clusters simultaneously have zero factors.
• Fixed storage bug in IM(I)FA models when learn.d is TRUE but learn.alpha is FALSE.
• Fixed density plot for discount when mutation rate is too low (i.e. too many zeros).
• Fixed simulation of loadings matrices for empty MIFA/OMIFA/IMIFA clusters using byrow=TRUE:
  Loop to simulate loadings matrices now generally faster also for all models.
• Fixed silly error re: way in which (I)FA models are treated as 1-cluster models to ensure they run:
  Related bug fixed for OM(I)FA/IM(I)FA models when starting number of clusters is actually supplied.

IMIFA v1.2.1 - (3^rd release [patch update]: 2017-05-29)

Improvements

• Posterior mean scores can now also be plotted in the form of a heat map (previously loadings only).
  load.meth argument replaced by logical heat.map in plot.Results_IMIFA.
• mat2cols gains compare argument to yield common palettes/breaks for heat maps of multiple matrices:
  Associated plot_cols function also fixed, and now unhidden.
• Removed certain dependencies with faster personal code: e.g. Procrustes rotation now quicker:
  IMIFA no longer depends on the corpcor, gclus, MASS, matrixcalc, or MCMCpack libraries.

Bug Fixes

• Used par()$bg (i.e. default "white") for plotting zero-valued entries of similarity matrix.
• Range of data for labelling in heat_legend calculated correctly.
• mcmc_IMIFA's verbose argument now governs printing of message & cat calls, but not stop or warning.
• Fixed storage and plotting of loadings, particularly when some but not all clusters have zero factors.
• Added NEWS.md to build.

IMIFA v1.2.0 - (2^nd release [minor update]: 2017-05-09)

New Features

• Learning the Pitman-Yor discount & alpha parameters via Metropolis-Hastings now implemented.
  • Spike-and-slab prior specified for discount: size of spike controlled by arg. kappa.
  • Plotting function's param argument gains the option discount for posterior inference.
• Sped up simulating cluster labels from unnormalised log probabilities using the Gumbel-Max trick (Yellott, 1977):
  gumbel_max replaces earlier function to sample cluster labels and is now unhidden/exported/documented.
• Added new plot when plot.meth=GQ for OM(I)FA/IM(I)FA models depicting trace of #s of active/non-empty clusters.
• Added function Zsimilarity to summarise posterior clustering by the sampled labels with minimum
  squared distance to a sparse similarity matrix constructed by averaging the adjacency matrices:
  when optionally called inside get_IMIFA_results, the similarity matrix can be plotted via plot.meth="zlabels".

Improvements

• Metropolis-Hastings updates implemented for alpha when discount is non-zero, rather than usual Gibbs.
  Mutation rate monitored rather than acceptance rate for Metropolis-Hastings updates of discount parameter.
• Fixed calculation of # 'free' parameters for aic.mcmc & bic.mcmc criteria when uniquenesses are isotropic:
  PGMM_dfree, which calculates # 'free' parameters for finite factor analytic mixture models is exported/documented.
  This function is also used to add checks on the Dirichlet hyperparameter for OM(I)FA methods.
• DIC model selection criterion now also available for infinite factor models (previously finite only).
• G_priorDensity now better reflects discrete nature of the density, and plots for non-zero PY discount values.
• Posterior mean loadings heatmaps now also display a colour key legend via new function heat_legend.
• Avoided redundant simulation of stick-breaking/mixing proportions under both types of IM(I)FA slice sampler.
• Simulated (finite) mixing proportions w/ Gamma(alpha, 1) trick (Devroye 1986, p.594) instead of MCMCpack:rdirichlet:
  rDirichlet replaces earlier function to sample mixing proportions and is now unhidden/exported/documented.
• Deferred setting dimnames attributes in mcmc_IMIFA to get_IMIFA_results: lower memory burden/faster simulations.
• Jettisoned superfluous duplicate material in object outputted from get_IMIFA_results to reduce size/simplify access.
• Restored the IMFA/IMIFA arg. trunc.G, the max allowable # active clusters, and # active clusters now stored.
• Code sped up when active G=1 by not simulating labels for IM(I)FA models.
• Reduced chance of crash by exceeding memory capacity; score.switch defaults to FALSE if # models ran is large.

Bug Fixes

• 2^nd IM(I)FA label switching move sped up/properly weighted to ensure uniform sampling of neighbouring cluster pairs.
• Offline label switching square assignment correction now permutes properly.
• Fixed factor score trace plots by extracting indices of stored samples using Rfast::sort_unique and rotating properly.
• Fixed adding of rnorm columns to scores matrix during adaptation, esp. when widest loadings matrix grows/shrinks.
• Fixed initialisation (and upper limit) of number of clusters for OM(I)FA/IM(I)FA, esp. when N < P.
• Updates of DP/PY alpha parameter now correctly depend on current # non-empty rather than active clusters.
• Fixed density plots for parameters with bounded support, accounting for spike at zero for discount.
• Slightly rearranged order Gibbs updates take place, esp. to ensure means enter simulation of uniquenesses properly.
• Edited/robustified subsetting of large objects when storing mcmc_IMIFA output.
• Tightened controls for when certain parameters are not stored for posterior inference.
• Edited Ledermann upper bound stop(...) for finite factor models to warning(...).
• Geometric rather than arithmetic mean used to derive single rate hyperparameter for PPCA's isotropic uniquenesses.
• Uniquenesses now stored correctly for all clustering methods.
• Indices of uncertain obs. returned (get_IMIFA_results)/printed (plot.Results_IMIFA) even when zlabels not supplied.
• Fixed behaviour of progress bar when verbose=FALSE.
• Fixed typos and expanded/clarified help documentation/vignette.

IMIFA v1.1.0 - (1^st release: 2017-02-02)

Reference manual