I have a dataset with large number of features and about 15 000 observations. I’m using a probability distribution distance metric related to Jensen-Shannon divergence (JSD) to cluster the observations calculated as described in http://enterotype.embl.de/enterotypes.html. I’m applying R implementation of Partitioning Around Medoids (PAM) clustering to my JSD distance matrix.
The issue is that the size of the distance matrix seems to be too big and eats all the memory. I’m looking for alternative implementations. k-means doesn’t work with other distance metrics than eucleidean, R clara works only with eucleidean and manhattan distance matrices. Sparks do not support pam or clara yet.
- Will this proposal https://stats.stackexchange.com/a/12503/46598 work for transforming my JSD distance matrix to eucleidean distance matrix?
- What else would you consider?
UPDATE 5.2.2015: Thanks, Aleksandr for thorough collection of links! I would actually like to stay in the scope of my question. I'm interested in similarities of my documents. The JSD distances express the document similarities very well and I'm happy with it. I would like to continue from that. I need to figure out for each document the most similar ones that are relevant from business point of view.
There are two overly naive approaches by ordering the similar documents for each document according to JSD distance and either select n first or n first below some JSD threshold as the similar ones.
The better approach is to let clustering algorithm to form the clusters because both of those naive approaches either cut off too many similar ones or identify similars also those that are actually not.
So my question is: which clustering algorithm is the most reasonable to use given that I have very good JSD distance matrix to feed in already.