README.md

In btaschler/mcap: Model-Based Clustering in High Dimensions via Adaptive Projections

MCAP

mcap provides a model-based clustering approach in very high dimensions (especially when p is much larger than n) via adaptive projections. Clustering is based on full variances Gaussian mixture modelling in a lower dimensional (projected) space. The projection dimension is set adaptively in a data-driven manner based on a cluster stability criterion. Available projection variants (so far) include PCA and random Projections (Gaussian as well as sparse methods).

Resources

See our paper: currently under review

preprint: …

Getting Started

Clone or download the code from github.

Alternatively, you can install mcap directly from github with:

# install.packages("devtools")
devtools::install_github("btaschler/mcap")

Prerequisites

Dependencies on other packages:

• for parallelisation: foreach, doParallel, parallel

• for clustering: pcaMethods, nethet, mclust, RandPro, kernlab

• misc: iterators, magrittr, stats, dplyr, tidyverse, utils, methods, data.table, RevoUtilsMath

Quick demo

This is a basic example showing how to use mcap to cluster two (known) groups:

library(mcap)

### basic example code
K <- 2       #number of clusters (groups)
n_k <- 200   #number of samples per group
p <- 1000    #number of features (dimension)
A <- matrix(rnorm(n_k*p), n_k, p)            #data for group 1
B <- matrix(rnorm(n_k*p, mean = 1), n_k, p)  #data for group 2
X <- rbind(A, B)                   #input matrix
Y <- c(rep(0, n_k), rep(1, n_k))   #known labels

## using PCA projections
model_fit <- MCAPfit(X, k = K, projection = 'PCA', centering_per_group = FALSE,
                     true_labels = Y, parallel = TRUE)

## sparse random projection
model_fit <- MCAPfit(X, k = K, projection = 'li', centering_per_group = FALSE,
                     true_labels = Y, parallel = TRUE)

## adjusted Rand index
print(model_fit$fit_gmm$aRI)

## display assigned cluster labels for each sample
print(model_fit$fit_gmm$model_fit$comp)

## show optimised projection dimension
print(model_fit$fit_q_opt$q_opt)

Versioning

For all available versions, see releases. We use Semantic Versioning.

Authors

• Bernd Taschler, Sach Mukherjee

List of contributors.

License

This project is licensed under the GNU General Public License – see the GPL-3.0 for details.

Acknowledgments

• Konstantinos Perrakis for valuable discussions.

• The coffee machine for mental and physical support.

btaschler/mcap documentation built on May 26, 2019, 1:31 a.m.