Kmeansplusplus

Kmeans++ in Julia

Adrian Haldenby a.haldenby@gmail.com

Overview

A simple (read sloppy) implementation of Kmeans++ in Julia. Algorithm that intelligently chooses the initial cluster centres before running Lloyd's algorithm. Check out https://normaldeviate.wordpress.com/2012/09/30/the-remarkable-k-means/ for more info.

This is was mostly an exercise to get me acquainted with Julia so definitely not for production.

Usage

Simply source kmeanspp.jl and run the algorithm with the following function call

output = run_kmeans(data,nclust,plusplus=true)

• data is an NxP matrix of data points
• nclust is an intger that specifies the number of clusters
• plusplus is a bool that's true if we want to use the ++ algorithm to initialize the cluster centers

Outputs a two item Dict of "centres" and the assigned "groups" for each of the points

test_kpp.jl contains a helper function to create some synthetic data from sampling a user user specified series of multivariate normal distributions with

test_data = generate_cluster_data(num_clusters,num_dims,num_entries)'

• num_clusters are the number of distributions to create
• num_dims is the number of dimensions the data will have (P)
• num_entries is the number of samples from each gaussian

Output is is an P+1xN matrix where the final row is are labels for each point