Hierarchial Clustering Methods

[rcurtin]

Reported by abhinav on 19 Dec 44926540 22:50 UTC
Sometimes, when we cannot decide how many clusters to choose, we can use Hierarchical Clustering Methods, as these are tailor made for this kind of clustering.
I have a month long vacation, and thinking of implementing this feature myself according to mlpack coding conventions.

Inspired by #370.

Migrated-From: http://trac.research.cc.gatech.edu/fastlab/ticket/374

[rcurtin]

Commented by rcurtin on 30 Nov 44927420 05:34 UTC
Sounds great to me; if you write it and have good tests for it I'll happily merge it into trunk. I wish I had a one month vacation... :)

[rcurtin]

Commented by abhinav on 8 Aug 44977005 17:03 UTC
I completed a simple naive O(n^3^) implementation. I am now going to improve it to O(n^2^ * log(n)) using Priority Queues. Presently, I am implementing Single, and Complete Hierarchical Clustering after which I would go for Centroid, Group-Average, Ward, etc.

[rcurtin]

Commented by rcurtin on 19 Apr 44977141 04:26 UTC
Sounds good to me. It's worth pointing out that single-linkage clustering is already implemented under the guise of "Euclidean Minimum Spanning Trees": see src/mlpack/methods/emst/ or the emst program. The process of finding a Euclidean minimum spanning tree is the same as the process for single-linkage hierarchical clustering. Probably a bit of modification needs to be done to make that code work for hierarchical clustering, but it should not be too hard.