Variational Mixture Models for Clustering Categorical Data

A variational Bayesian finite mixture model for the clustering of categorical data, and can implement variable selection and semi-supervised outcome guiding if desired. Incorporates an option to perform model averaging over multiple initialisations to reduce the effects of local optima and improve the automatic estimation of the true number of clusters. For further details, see the paper by Rao and Kirk (2024) .

VICatMix: Variational Inference for Categorical Mixture Models

VICatMix is a variational Bayesian finite mixture model designed for the clustering of categorical data, implemented as an R package incorporating C++ (via Rcpp and RcppArmadillo) for faster computation. The package provides options to include variable selection to enhance its performance on high-dimensional or noisy data, and to incorporate model averaging and summarisation over multiple different initialisations for improved accuracy. The package additionally contains functions to generate sample clustered binary/categorical data for testing. For more details on the model, please refer to the arXiv preprint.

Installation

To install the VICatMix package, you can use the devtools package to install directly from GitHub:

install.packages("devtools")
devtools::install_github("j-ackierao/VICatMix")
library(VICatMix)

Note VICatMix depends on the Rcpp and RcppArmadillo packages, which both require an appropriate C++ compiler.

Examples

An example of generating sample binary data for clustering, with 'true' cluster labels.

generatedData <- generateSampleDataBin(1000, 4, c(0.1, 0.2, 0.3, 0.4), 100, 0)

An example of running one initialisation of VICatMix on sample data without variable selection:

result <- runVICatMix(generatedData[[1]], 10, 0.01) 

An example of implementing model averaging over 30 initialisations of VICatMix on sample data with variable selection:

result <- runVICatMixVarSelAvg(generatedData[[1]], 10, 0.01, inits = 30)