4
$\begingroup$

I came across a variant of k-means that adds a algorithm to select good starting values called k-means++.

kMeansInitializer[data_, k_Integer] :=
 Module[
  {startingPoint = RandomChoice[data], getDistance, getDistances, 
   nextPoint},
  getDistance[datum_, points_] := Min[Norm[datum - #]^2 & /@ points];
  getDistances[points_] := getDistance[#, points] & /@ data;
  nextPoint[points_] := RandomChoice[getDistances[points] -> data];
  NestList[nextPoint, startingPoint, k - 1]
  ]

After implementing the algorithm (a bit of a challenge for someone at my level) I now find I have no idea how to feed the values I generate to ClusteringComponents to use as initial values for use in its k-means implementation.

$\endgroup$
  • $\begingroup$ I don't think it is possible to do that via ClusteringComponents. Here's an implementation of k-means that I wrote, which might be easily extended to your needs. $\endgroup$ – rm -rf Oct 3 '12 at 23:33
0
$\begingroup$

Posting rm -rf's comment as an answer for the purpose of improving the stats of this site, it is not possible to provide ClusteringComponents with initial values. It is almost certainly already implemented using an algorithm to choose good starting points, and probably because they all converge to the same solution it was not deemed necessary to make it possible to change the starting points.

But as rm -rf said he has implemented the most common algorithm for k-means and made it available here. You can provide your own starting points by changing

m = RandomSample[list, k]

to

m = kMeansInitializer[list, k]
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.