R/clustercheck.R

In CAMCOS/camcos2017: CAMCOS Spring 2017: Efficient and Scalable Spectral Clustering for Document Clustering.

#' Checking cluster accuracy
#'
#' In the results table, true clusters are the rows, guessed clusters are the
#' columns
#'
#' @param cluster Integer vector. Labels from clustering
#' @param Labels Integer vector. Ground truth labels.
#' @param k Integer. Number of clusters.
#'
#' @return List of 3. Confusion matrix, label mapping, accuracy.
#' @export
#'
#' @examples
#' # Example of a results matrix that FAILS clustercheck
#' # Cluster assignment of "true cluster #3" is split too evenly
#' results <- matrix(c(15,31,4,452,5,0,1,384,1,20,1,3,34,141,56,53,289,357,0,
#'   2,282,6,160,5,527,32,48,51,9,6,4,2,203,0,0,5), nrow = 6)
clustercheck <- function (cluster, Labels, k) {


  comp <- cbind(cluster, Labels) # nx2 matrix of cluster assignments and true group membership

  # results matrix

  results <- matrix( rep(0,k^2), nrow = k)


  # results matrix: rows = true group, columns = assigned cluster

  for (row in 1:nrow(comp)){

      results[comp[row,2], comp[row,1]] <- results[comp[row,2], comp[row,1]] + 1

  }


  # calculate percentage distribution of each true group

  percent.cluster <- t(apply(results, 1, function(x) x/sum(x)))

  assignment <- matrix( rep(0,k), nrow = k)


  # assign the most-correct true group to that cluster, and continue

  for (index in 1:k) {

      test <- which(percent.cluster == max(percent.cluster), arr.ind = TRUE)

      assignment[ test[1],1 ] <- test[2]

      percent.cluster[test[1],] <- 0   # remove that row (true group) from consideration

      percent.cluster[,test[2]] <- 0   # remove that column (cluster) from consideration

  }


  # calculate percentage assigned

  correct <- 0
  for (i in 1:nrow(assignment)){

      correct <- correct + results[i,assignment[i,1]]

  }

  percent.score <- correct / sum(results)


  # clean up the assignment matrix

  assignment <- cbind( seq(1,k), assignment[,1])
  colnames(assignment) <- c("True Cluster", "Assigned Cluster")


  return(list(results, assignment, percent.score))

}