0
$\begingroup$

I'm new to clustering. recently I've implemented agglomeration clustering, I understood that we merge clusters that have distance < Threshold.

  1. But how to fix the Threshold?
  2. How to choose right method for finding distance [euclidean, manhattan,..]?
  3. How to evaluate my clusters?
$\endgroup$
1
  • $\begingroup$ Points 2 and 3 are so capacious that nobody will consider answering them. If you are new to clustering, please read books, or at least one, first. Point 1 is not clear - what you mean by threshold; aglomeration clustering algorithm does not use "thresholds". $\endgroup$
    – ttnphns
    Sep 25, 2013 at 17:19

1 Answer 1

1
$\begingroup$
  1. There are different strategies for extracting flat clusters. Thresholds are good when you have an intuition, e.g. "1 mile". Alternatively, you can choose the cut to get a desired number of clusters.

    The following publication may also be of use:

    Ricardo J. G. B. Campello, Davoud Moulavi, Arthur Zimek, Jörg Sander:
    A framework for semi-supervised and unsupervised optimal extraction of clusters from hierarchies
    Data Mining and Knowledge Discovery, Volume 27, Number 3 / November 2013.

  2. Choose whichever distance function works best for your data. Spend at least as much time on preprocessing as on choosing the distance. There are no general rules of which is best. I have good experiences with using Canberra as default.

  3. Read a book on this. There are way too many measures. Adjusted Rand Index is a good default choice, I guess. When using internal evaluation, make sure it is not closely related to your algorithms, because you can then overfit on the method side. E.g. K-means usually comes out best when you look at variances...

$\endgroup$
1
  • $\begingroup$ Thanks for a reference to the article, it can be found in the web. $\endgroup$
    – ttnphns
    Sep 25, 2013 at 20:25

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.