batchmix (version 2.2.1)

minVI: Minimium VI

Description

Local implementation of S. Wade's `minVI` function from their `mcclust.ext` package (available from github). Reimplemented here to avoid dependency on a non-CRAN package and we have dropped the `greedy` method. Finds the optimal partition by minimising the lower bound to the Variation of Information obtained from Jensen's inequality where the expectation and log are reversed. For full details please see the aforementioned package and Wade and Ghahramani, 2018, 'Bayesian Cluster Analysis: Point Estimation and Credible Balls (with Discussion)'.

Usage

minVI(psm, cls.draw = NULL, method = "avg", max.k = NULL)

Value

If `method` is `'all'` returns a matrix of four clusterings, one for each method and a repeat of that which performs best based on minimising the Variation of Information between the clustering and the PSM. Otherwise returns a vector. This is annotated with the attribute `"info"`, a named list describing:

* `.$loss`: the loss score used (Variation of Information)

* `.$maxNClusters`: the `max.k` value used by the `'comp'` and `'avg'` methods

* `.$expectedLoss`: the estimated minimum Variation of Information for the point clustering(s)

* `.$method`: the point method used to infer the clustering(s)

Names are due to legacy reasons - this function is replacing the `salso::salso` function and name choices are to minimise workflow damage.

Arguments

psm

The posterior similarity matrix for a set of clustering MCMC samples such as is returned by the `createSimilarityMat` function.

cls.draw

The set of clustering MCMC samples used to generate `psm`. Only required if `method` is one of `'draws'` or `'all'`.

method

String indicating which method is used to find the point estimate clustering. Must be one of `'avg'`, `'comp'`, `'draws'` or `'all'`. Defaults to `'avg'`. If `'all'` is passed the three methods are all applied to return different choices of point clustering.

max.k

The maximum number of clusters to consider. Only used by the `'comp'` and `'avg'` methods. Defaults to one-quarter the number of data points rounded up.

Examples

if (FALSE) {
# MCMC samples and BIC vector
mcmc_outputs <- runMCMCChains(
  X,
  n_chains,
  R,
  thin,
  batch_vec,
  type
)

# Note that in this toy example we have not applied a burn in
psm <- createSimilarityMat(mcmc_outputs[[1]]$samples)
cl_est <- minVI(psm, mcmc_outputs[[1]]$samples)
}