1
$\begingroup$

I have n articles and a list of articles for each article implying similarity, e.g., a1 -> a3, a5 means a3 & a5 are similar to a1. Similarity here is 0 or 1 that is either two articles are similar or not; there is no floating point value.

Now, I want to cluster these articles into let's say k clusters. I can't directly apply KMeans to it, since these articles are not in vector form. Also I could try Spectral Clustering, but I don't understand in what form output is returned by Spectral Clustering (or Power Iteration Clustering).

$\endgroup$
1
  • 1
    $\begingroup$ Your data are presently in the format of adjacency list. But at any time it could be converted into the square matrix of pairwise similarity - if a program you are going to use requests it. But the question is another one. Your similarity values are just binary, yes and no. In other words, the graph is non-weighted. No clustering is really needed in your case (and K-means especially!) since clustering implies quantitative similarity (weighted graph). You should try some of many plotting techniques (including various networks) for non-weighted graphs. $\endgroup$
    – ttnphns
    Aug 15 '15 at 11:39
1
$\begingroup$

Also consider:

  • DBSCAN
  • Single-link (i.e. connected components)
  • Complete-link (i.e. Cliques)
  • Average-link (near cliques)

Plus, all the literature you can find on *graph clustering|.

$\endgroup$

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.