0

I have a large data set with BINARY user/items feature matrix:

  1. I need to cluster both users and items. Is there anyway to do them simultaneously in Mahout?
  2. More importantly, if I use loglikelihood as a similarity measure, what clustering algorithms will actually support such distance metric to cluster the data?
Peter O.
  • 32,158
  • 14
  • 82
  • 96
user1848018
  • 1,086
  • 1
  • 14
  • 33

1 Answers1

1

No, clustering by users and items are separate processes. Though in spirit it's exactly the same process, just applied two different ways.

If you want more specific answers within Mahout you'd have to say more about what parts of the code you are using because there are several different parts that involve clustering.

There are some agglomerative clustering pieces in the project, which works for any similarity metric. The other implementations that I know of are definitely of the "k-means" variety, assuming a continuous vector space and not vectors over {0,1}. You would need a k-medoids kind of algorithm I think and this isn't in the project that I know of.

Sean Owen
  • 66,182
  • 23
  • 141
  • 173
  • Thanks Sean, Could you please be more specific about the agglomerative clustering parts in Mahout. I am in the design stage and need to know if I have the similarity matrix, what would be a good clustering algorithm to use based on this measure to cluster the data. – user1848018 Nov 23 '12 at 20:43
  • I am thinking of `TreeClusteringRecommender`, which is old and non-distributed code that I made, and not something I'd particularly recommend to anyone. But as it is not centroid-based you only need a similarity metric. In general the answer to your problem is 'k-medoids'. – Sean Owen Nov 23 '12 at 21:54