Talk:Single-linkage clustering

Ex aequos
I disagree with the statement that

implementing a different linkage is simply a matter of using a different formula to calculate inter-cluster distances in the initial computation of the proximity matrix

With single linkage, if there are ex-aequos, the order in which they are considered does not matter. But with complete linkage, the order does matter, and may change the number of clusters.

Consider the set {1,2,3,4} with distance(x,y) = abs(x-y). The minimal distance is between 1 and 2, 2 and 3, 3 and 4. If we start with {2,3}, we end up with 3 clusters at level 1: {1}, {2,3}, {4}, because 1 and 4 are now at distance 2 of the central cluster. On the other hand, if we start with {1,2}, we can still cluster 3 with 4 and have only 2 clusters at level 1: {1,2}, {3,4}.

So complete linkage clustering requires a more educated choice at step 2, in order to be deterministic. Being not a specialist in hierarchical clustering, I have no answer to this problem... As a matter of fact, I was looking for the answer in Wikipedia :)

--Pchampin (talk) 22:15, 24 February 2009 (UTC)


 * What do you think of the clarification I added? --Pot (talk) 12:47, 25 February 2009 (UTC)
 * What clarification did you mean? The mistake pointed by Pchampin is still here.37.110.18.166 (talk) 10:05, 7 March 2014 (UTC)

Added references
I added a couple references and a little text. I don't have the time/skill right now to give details of SLINK without violating copyright, but they are in Gan et al.. If someone wants to give more detail, that would be great. PeterLFlomPhD (talk) 14:02, 27 July 2015 (UTC)