1
$\begingroup$

I am trying to cluster 6 PC scores for a 1600 cases data (so its a 1600 x 6 matrix). I am using average linkage clustering technique that would be followed by a K mean clustering. The purpose of Average linkage clustering is to provide seed values for K mean clustering.

Would somebody explain me how to extract the results of average linkage clustering from Agglomeration Schedule?

$\endgroup$
2
$\begingroup$

You need not to use agglomeration schedule for your task. You need to (1) decide how many clusters are there, (2) save this cluster solution as cluster membership variable, (3) compute centres - vector of means, for each cluster, (4) input it into K-MEANS clustering as "initial centres".

Perhaps you ought to do (2) first (save range of solutions) and (1) then (decide upon the "best" solution with the advise of some clustering criterion or other approach).

Note, however, that 1600 cases is too much for hierarchical clustering - not just because of computer-performance issues, but because hierarchical clustering is one-path greedy algorithm and therefore is prone to produce suboptimal results on distant steps of agglomeration.

$\endgroup$
  • $\begingroup$ Can you please explain how to do the 4 steps in SPSS step by step? $\endgroup$ – mzalikhan Jun 30 '11 at 2:57
  • 1
    $\begingroup$ I won't explain point 1 because it is art. For point 2, use button Save in Hier. Cluster procedure. For point 3, use Aggregate procedure. For point 4, use check Read initial in K Means procedure (to know your dataset with input centres must look like, first save centres with Write final and glance at the output dataset). $\endgroup$ – ttnphns Jun 30 '11 at 5:20
1
$\begingroup$

You don't say why you chose the average linkage method, but since you are only doing that step to generate starting values for k means, the choice of method in the first step might not be very important. Ward's method, also available in the SPSS CLUSTER procedure, may scale better to larger datasets. If the dataset for the first stage turns out to be too large, you could take a random sample to calculate the initial cluster centers.

HTH, Jon Peck

$\endgroup$
  • $\begingroup$ If you have experience of calculating initial centers by taking random samples, please do let me know how big sample is good enough for this purpose? and also if there is any special technique to take the random sample? $\endgroup$ – mzalikhan Jul 1 '11 at 2:59

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.