kmeans 1 d data

I have data = [1, 1, 2, 3, 10, 11, 13, 13, 17]

I want to to cluster them in 3 clusters..like (1,1,2,3) in cluster 1 (10,11) in cluster 2 and (13,13,17) in cluster 3.

but when I apply idx = kmeans( data, 3 ); kmeans clusters them with random cluster index..

Is there any way to fix that?

How could I plot my clustered data ? scatter doesn't work here

[Geoff Hayes]

Tusu - please clarify what you mean by kmeans clusters them with random cluster index. Do you mean that sometimes the algorithm returns the clusters that you have defined above and other times the algorithm returns some other cluster pattern? You could try to specify the centroid starting locations of the clusters. Something like

 kmeans(data,3,'Start',[2 10 13])

[tusu]

No cluster pattern remains the same but some time say (1,1,2,3) belongs to cluster 1 some time its belongs cluster 2..I want kmeans should allocate (1,1,2,3) in cluster 1, (10,11) in cluster 2 and (13,13,17) in cluster 3 in fixed manner.

Answer 1

[Image Analyst]

If you know the boundaries that you want in advance , then there is no reason to use kmeans . Just classify them yourself

for k = 1 length(data)
    if data(k) <= 3
        theClasses(k) = 1;
    elseif data(k) < 13
        theClasses(k) = 2;
    else
        theClasses(k) = 3;
    end
end

[tusu]

Thanks for response. I actually don't know the boundaries..Kmeans would select the boundary..as this data set is sorted array...I need smaller value should be in 1st cluster and larger value shall be in higher index cluster (say cluster 3) can it be done? How this 1d clustering result can be plotted?

[Image Analyst]

If you don't know the boundaries, then how can you say what kmeans is doing is wrong? Give some examples of how kmeans is clustering them so we can see that it's just giving random classifications as you claimed.

[Tom Lane]

For one-dimensional data you might just sort the cluster means, and re-label the cluster numbers you got to match the sort order. The cluster numbers are just names, and aren't intended to have any real meaning.

[Image Analyst]

Tusu:

If you want to relabel the cluster labels according to something, like for example the magnitude of the cluster centroid, see my attached demo. So,if your clusters are initially labeled

[1, 1, 2, 3, 10, 11, 13, 13, 17] % Data

[2, 2, 2, 2, 3, 3, 1, 1, 1] % Initial labels caused by randomness in kmeans() alrogithm

you can use my attached code to relabel them like

[1, 1, 2, 3, 10, 11, 13, 13, 17] % Data

[1, 1, 1, 1, 2, 2, 3, 3, 3] % Final labels based on the mean of each cluster

Demo code (you might need to run it several times to get the initial order to be NOT sequential):

data = [1, 1, 2, 3, 10, 11, 13, 13, 17]
[classLabels, centroids] = kmeans(data', 3);
classLabels' % Show in command window.
centroids' % Show in command window.
% Sort the centroids by distance from 0
[sortedCentroid, sortOrder] = sort(centroids, 'ascend')
% Instantiate new labels
newClassLabels = zeros(size(classLabels));
% Now relabel each point with the class label that we WANT it to be.
for k = 1 : length(classLabels)
	originalLabel = classLabels(k);
	newClassLabels(k) = find(sortOrder == originalLabel);
	fprintf('For point #%d (= %2d), an original class label of %d will now be %d\n', k, data(k), originalLabel,newClassLabels(k));
end

Answer 2

[Javier Buldu]

data=data'

Then:

idx = kmeans(data, 3 );