Small anticlust simulation

small_anticlust_simulation.R

# Test if splitting data via anticlustering leads to closer groups means to the *true* population means,
# as compared to a random split (e.g., for cross validation
simulate <- function(N = 100, split = c(1, 3) / 4) { # default: split 75/25
  data <- rnorm(N)
  groups <- anticlustering(
    data,
    K = round(N * split),
    objective = "variance"
  )
  c(
    anticlust = total_deviation(data, groups),
    random = total_deviation(data, sample(groups))
  )
}

total_deviation <- function(data, groups, squared = FALSE) {
  if (squared) {
    return(sum(tapply(data, groups, mean)^2))
  }
  sum(tapply(data, groups, mean)^2)
}

rowMeans(replicate(500, simulate()))
#> anticlust    random 
#> 0.1558745 0.2424823 

rowMeans(replicate(500, simulate(squared = TRUE)))
#>  anticlust     random 
#> 0.02033793 0.05104909 
# anticlustering has lower deviation from *true* mean