2
$\begingroup$

In Spectral Clustering, the algorithm suggests performing K-means to $k$ eigenvectors of the resulted Laplacian matrix.

Can I use other clustering algorithms such as k-medoids or other non-distance based algorithms instead of k-means, or is the algorithm designed to best respond to k-means?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

The point is to find the location of the zeros in the eigenvectors. You could use PCA or any other form of clustering.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you for your response. There is a strong correlation between my data points such that the correlation matrix (cor(X') has a lot of elements larger than 0.9. I set any value greater 0,9 to zero to make the matrix sparse. Do you agree with that? $\endgroup$
    – H_A
    Commented Jul 28, 2014 at 19:53
  • $\begingroup$ Why would random variables with a correlation of greater than 0.9 be forced to have a correlation of 0? When I said you could use any clustering technique, I meant it was on the eigenvectors of the Laplacian matrix. It has been seen that using different techniques produce slightly different answers. My advice is to not include personal heuristics as far as possible (i.e. the data says what it says) $\endgroup$
    – Sid
    Commented Jul 28, 2014 at 20:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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