Low Memory Use Trimmed K-Means

Performs the trimmed k-means clustering algorithm with lower memory use. It also provides a number of utility functions such as BIC calculations.

lowmemtkmeans

The tkmeans package attempts to implement the trimmed k-means algorithm of Garcia-Escudero, et. al.(2008) using as little memory as possible. Data is edited in place, the trimming is implemented using a priority queue structure in C++ through Rcpp and low memory use versions of utility functions are provided.

An extremely simple example:

1. Convert the iris dataset to a matrix and rescale matrix columns.

   iris_mat <- as.matrix(iris[,1:4]) scale_params<-scale_mat_inplace(iris_mat)

2. Cluster with 2 and 3 clusters, 10% trimming

iris_cluster_2<- tkmeans(iris_mat, 2 , 0.1, c(1,1,1,1), 1, 10, 0.001)  
iris_cluster_3<- tkmeans(iris_mat, 2 , 0.1, c(1,1,1,1), 1, 10, 0.001)

1. Calculate BIC

BIC_2 <-cluster_BIC(iris_mat, iris_cluster_2)  
BIC_3 <-cluster_BIC(iris_mat, iris_cluster_3)

1. Allocate using 3 clustering

clustering <- nearest_cluster(iris_mat, iris_cluster_3)

1. Plot results using reconstructed matrix

library(lattice) 
orig_matrix <- sweep(sweep(m,2,scale_params[2,],'*'),2,scale_params [1,], '+')  
xyplot(orig_matrix[,1]~orig_matrix[,2], group=clustering) 

To install the latest version:

install.packages("devtools")  
library(devtools)  
install_github("andrewthomasjones/tkmeans")