k-Means ++ center initialization algorithm

kMeansPP.R

require("proxy")
library(proxy)


# kmeans++ center initialization algorithm
kMeansPP <- function(df, k, doPlot = TRUE){
  kCenters <- data.frame(matrix(NA, ncol = ncol(df), nrow = k))
  whichPoints <- rep(NA, k)
  whichPoints[1] <- sample(1:nrow(df), 1)
  kCenters[1, ] <- df[whichPoints[1], ]  # Initial center
  
  for(kk in 2:k){
    distMat <- proxy::dist(df, kCenters[1:(kk-1), ])
    distToNearestCenter <- apply(distMat, 1, min)
    whichPoints[kk] <- sample(1:nrow(df), 1, prob = distToNearestCenter^2)
    kCenters[kk, ] <- df[whichPoints[kk], ]
  }
  
  if(doPlot == TRUE){
    plot(df[, 1:2], col = "GRAY")
    points(kCenters[, 1:2], col = 1:k, pch = 20)
  }
  
  outList <- NULL
  outList$Centers <- kCenters
  outList$whichPoints <- whichPoints
  return(outList) 
}

sample kMeansPP.R

# Test it
myData <- data.frame(x = rnorm(100), y = rnorm(100), z = rnorm(100))
plot(myData)
 
PPresult <- kMeansPP(myData, 3)
 
convergedClusters <- kmeans(myData, centers = PPresult$Centers)
points(convergedClusters$centers, col = 1:k, pch = 10)
points(myData, col = convergedClusters$cluster, cex = 1/2)