I am trying to use clustering on certain data. The data itself has three natural levels: at the lowest level the elements are fundamental building blocks, at the second level these fundamental building blocks merge together to become larger element, and then at an even higher level they become largest elements. Analogically, they may be think of small particles, atoms and matter. This is just analogy.

I want to do hierarchical clustering for the lowest levels using the raw features. Then, when the threshold hits the point at the middle level (atoms), I want to start clustering them in terms of some aggregate features. And similarly, when I hit the third natural level (matter), I want to start clustering them using further aggregation of features.

Is there any specific variant of hierarchical clustering that is suitable for this purpose?


Edit: As it is pointed out in one of the comments, I am looking for an algorithm which performs hierarchical clustering while at certain levels can switch to different (aggregate of previous ones in some way) distance metrics.

  • 1
    $\begingroup$ What if the "clusters" at these levels are not consistent with your theory? I.e., you don't have a known atom or known matter? How would you compute these metrics? $\endgroup$ – Anony-Mousse Apr 28 at 8:36
  • $\begingroup$ @Aniny-Mousse, that is a good point. If at the clustering doesnt match with the 'natural' clustering, it may probably mean that I might not have athe complete list of features. Or perhaps the natural clusters should be questioned - in the atoms etc case, the natural clustering is the ground truth, but in other non scientific phenomena, it may challenge the 'natural' clustering. $\endgroup$ – dbm Apr 28 at 16:06

You can simply cluster at each level separately and combine the resulting trees into one.

For this, replace each leaf with its corresponding tree.

  • $\begingroup$ I am looking for more coherent method rather than joining different pieces, if available at all. $\endgroup$ – jjal Apr 27 at 14:47
  • $\begingroup$ What is incoherent here? It builds the result consistent with your "natural" levels. If you want something else, please describe that more clearly in your question. $\endgroup$ – Anony-Mousse Apr 27 at 17:32
  • 2
    $\begingroup$ I am looking for a method that takes care of the natural hierarchy of the data itself. The one you described is a trivial and obvious idea. If there is no such method, there is no such method. I am trying to confirm it either way here. I think this should be clear in the question. $\endgroup$ – jjal Apr 27 at 21:13
  • $\begingroup$ Define "takes care of the natural hierarchy". This is too vague. Doesn't my suggested approach do that? Why not? There is nothing wrong with a "trivial" approach if it satisfies the requirements. Do you want to switch distance measures inbetween? That will be hard if not enforcing exactly this composition. $\endgroup$ – Anony-Mousse Apr 28 at 2:42
  • $\begingroup$ Anony-Mousse, I have edited my question following your comment. Thanks. $\endgroup$ – jjal Apr 28 at 3:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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