# How to estimate K for K-means clustring

17 views (last 30 days)
wisekily on 15 May 2016
Commented: Bashar Saad on 12 Jul 2019
I'm working on unsupervised classification or clustering, i want to estimate the K (which refers to cluster number) before starting th k-means algorithm

Walter Roberson on 16 May 2016
You will probably not find any code already implemented for this purpose.
The theoretical answer for the "best" number of clusters to use is "one cluster for every unique point", as that will always have the best possible fit.
If you do not wish to use one cluster for every unique point, you need to have some kind of penalty term that favors fewer clusters. I read through the theory paper on that a few years ago, and it was clear to me that they were setting the weights arbitrarily (but usefully for the kinds of clustering they were doing), and that there was no way to calculate what the weights should be without some knowledge of the range of number of clusters that would be appropriate for the physical system being examined. The theoretical algorithms were not suitable for "unsupervised learning", only for "supervised learning". The work we were doing at the time required unsupervised learning, so there was no way for us to determine what the proper number of clusters should be.

the cyclist on 15 May 2016
This is not really a MATLAB question, but rather a general data science question.
Image Analyst on 15 May 2016
There are MATLAB functions for estimating the best k. I don't remember what they were - I'd have to look them up in the Machine Learning course notes.
wisekily on 15 May 2016

Image Analyst on 15 May 2016
The web page on kmeans explains how you can use silhouette() to determine the best number of clusters, k:
##### 3 CommentsShow 1 older commentHide 1 older comment
Walter Roberson on 16 May 2016
the cyclist on 16 May 2016
Which is also the same link that I pointed you to earlier. So, uh, now you have 3 of the top 10 contributors to this forum telling you consistently the same thing.

kira on 2 May 2019
old question, but I just found a way myself looking at matlab documentation:
klist=2:n;%the number of clusters you want to try
myfunc = @(X,K)(kmeans(X, K));
eva = evalclusters(net.IW{1},myfunc,'CalinskiHarabasz','klist',klist)
classes=kmeans(net.IW{1},eva.OptimalK);
Bashar Saad on 12 Jul 2019
could you help me pleas the code is not clear