Implementation of kmeans clustering using R

Question

I have implemented kmeans clustering on iris dataset (inbuilt dataset) in R. The code is given below:

X=as.matrix(iris[-5]);

K=3;

prevCentroids=matrix(0,K,dim(X)[2]);
centroids=X[sample(1:dim(X)[1],K),];

dot=numeric(3);
C=numeric(150);

while(!isTRUE(all.equal(centroids,prevCentroids)))
{
  for(i in 1:dim(X)[1])
   {
    for(j in 1:dim(centroids)[1])
     {
       dot[j]=(X[i,]-centroids[j,])%*%(X[i,]-centroids[j,]);
     }
     C[i]=which.min(dot);
   }

  prevCentroids=centroids;

  for(k in 1:K)
   {
     centroids[k,]=colMeans(X[which(C==k),]);
   }
}

print(cbind(iris,C));

Sometimes, with this code, I get 85% clustering correct. But sometimes, it is just 37% as correct, if I compare it with the already clustered inbuilt iris dataset.

Could anyone please tell me where I am going wrong?

Answer 1

Actually, this is the expected behavior for running a k-means algorithm on the iris dataset. The iris dataset contains only two distinct clusters. So if you randomly set up three centroids, the third centroid will either end up on the right or on the wrong cluster, causing the algorithm to split that cluster into two (see the left picture).

Your outcome with 37% accuracy will most probably look somewhat similar to the left picture where the left cluster has been split into two clusters. The 85% accuracy will most probably come from those times in which the right cluster has been split into two clusters (you cannot be sure though without printing the results).

For testing purposes, you can simply use set.seed(123) before running the algorithm, so that you always get the same results. However, this doesn't solve the problem that the iris dataset isn't ideal for clustering purposes. I would suggest you check out other datasets or you give the R-package cluster.datasets a try.