In psobczyk/public_varclust: Variable Clustering

Tutorial for varclust package

Introduction

varclust is a package that enables dimension reduction via variables clustering. We assume that each group of variables can be summarized with few latent variables.

This tutorial will gently introduce you to usage of package varclust and familiarize with its options.

You can install varclust from github (current development version).

install_github("psobczyk/varclust")

Main usage example

library(varclust)
library(mclust)

Let us consider some real genomic data. We're going to use FactoMineR package data. This data consists of two types of variables. First group are gene expression data. The second is RNA data. Please note that it may take few minutes to run the following code:

comp <- read.table("http://factominer.free.fr/docs/gene.csv",sep=";",header=T,row.names=1) 
benchmarkClustering <- c(rep(1, 68), rep(2, 356))    
comp <- data.frame(comp[,-ncol(comp)])   
set.seed(1)
mlcc.fit <- mlcc.bic(comp, numb.clusters = 1:15, numb.runs = 10, max.dim = 4, greedy = TRUE, 
                     estimate.dimensions = TRUE, numb.cores = 1, verbose = FALSE)
print(mlcc.fit)
plot(mlcc.fit)
mclust::adjustedRandIndex(mlcc.fit$segmentation, benchmarkClustering)

One might be a little concerned by high number of clusters found and that our benchmark clustering was not recognized, however, what we expect to see is the seperation of RNA and expression groups, not the homogenity within those groups. If we define different partion based on mlcc.fit we're getting very high Adjusted Rand Index.

part <- NULL
part[mlcc.fit$segmentation %in% c(4,5,10,12)] <- 1
part[!mlcc.fit$segmentation %in% c(4,5,10,12)] <- 2
mclust::adjustedRandIndex(part, benchmarkClustering)
misclassification(part, benchmarkClustering, max(table(benchmarkClustering)), 2)

If you know what is the true number of clusters you might use mlcc.reps function instead. In that case you can also compare segmentations in terms of misclassification.

mlcc.fit2 <- mlcc.reps(X=comp, numb.clusters = 2, numb.runs = 20, max.dim = 4, numb.cores=1)
mlcc.fit2
mclust::adjustedRandIndex(mlcc.fit2$segmentation, benchmarkClustering)
misclassification(mlcc.fit2$segmentation, benchmarkClustering, max(table(benchmarkClustering)), 2)

Running algorithm with some initial segmentation

You should also use mlcc.reps function if you have some apriori knowledge regarding true segmentation. You can enforce starting point

mlcc.fit3 <- mlcc.reps(comp, numb.clusters = 2, numb.runs = 0, max.dim = 2, 
                       initial.segmentations=list(benchmarkClustering), numb.cores=1)
mclust::adjustedRandIndex(mlcc.fit3$segmentation, benchmarkClustering)
misclassification(mlcc.fit3$segmentation, benchmarkClustering, max(table(benchmarkClustering)), 2)

Execution time

Execution time of mlcc.bic depends mainly on:

1. Number of clusters (numb.clusters)
2. Number of variables
3. Number of runs of k-means algorithm (numb.runs)

For a dataset of 1000 variables and 10 clusters computation takes about 8 minutes on Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz.

Choosing values of parameters

• If possible one should use multiple cores for computation. By default all but one cores are used. User can override this with numb.cores parameter
• For more precise segmentation one should increase numb.runs. Default value is 20
• Parameter max.dim should reflect how large we expect subspaces to be. Default value is 4
• If parameter greedy is TRUE (value set by default) the number of clusters is estimated in a greedy way. So program stops after getting first BIC local maximum
• If estimate.dimensions is TRUE subspaces dimensions are estimated. Otherwise all subspaces are assumed to be of dimension max.dim

psobczyk/public_varclust documentation built on May 26, 2019, 10:33 a.m.