K mean implementation: two implementations, same results, but different - but where is the difference?

Hi! I am supposed to do k-means implementation for a course. And I did it. My problem is: I implemented one part of it in two different ways, both deliver the same results und testsamples but only one passed the (automatical) test. And I have no idea why?

This routine is supposed to calculate the new centroids of the k-means algo. in the following code X is a Matrix of samples, idx is the centroids assignment for every sample for the current centroids and K is the number of centroids.

It starts with some basic values:

if true
  [m n] = size(X);
  centroids = zeros(K, n);
end

And this is my first and preferential solution:

if true
  for i=1:K
      XTemp = X(idx==i,:);
      centroids(i,:) = mean(XTemp(:,:));
  end
end

In the end each and every centroid has some new coordinates.

And this is the ugly implementation that finally did it:

if true
X_sum = zeros();
for i = 1:K
      X_sum = 0;
      counter = 0;
      for j = 1:m
              if i == idx(j)
                      X_sum = X_sum + X(j, :);
                      counter = counter +1;
              end
      end
      centroids(i, :) = X_sum/counter;
end
end

And now my question: What is the difference? - Because I don't see any!

Thanks.

Answer 1

[Walter Roberson]

mean(XTemp(:,:)) takes mean along the first dimension by default. The loop you programmed is taking the mean along the second dimension. mean(XTemp,2) should have the same effect.

[Andreas Kallmeyer]

OK. Thanks.