topic Re: Initial seed for clustering in SAS Procedures

Initial seed for clustering

ria_tran — Sat, 02 Jan 2016 21:23:41 GMT

Hi all,

Can you please explain how the initial seed can be obtained so that clustering can be repeated every run using the FASTCLUS and HPCLUS procedures (https://communities.sas.com/t5/SAS-Communities-Library/Tip-K-means-clustering-in-SAS-comparing-PROC-FASTCLUS-and-PROC/ta-p/221369) ?

Many thanks.

Re: Initial seed for clustering

PGStats — Sun, 03 Jan 2016 05:52:36 GMT

I couldn't follow the link you provided.

Try running a zero iteration clustering step to generate reusable seeds

proc fastclus data=sashelp.iris maxc=5 maxiter=0 outseed=seeds;
   var SepalLength SepalWidth PetalLength PetalWidth;
run;

and then reuse those seeds in subsequent clustering runs

proc fastclus data=sashelp.iris maxc=3 maxiter=100 seed=seeds;
   var SepalLength SepalWidth PetalLength PetalWidth;
run;

Re: Initial seed for clustering

ria_tran — Sun, 03 Jan 2016 14:23:52 GMT

Hi PGStats,

Thank you for your response. The link is: https://communities.sas.com/t5/SAS-Communities-Library/Tip-K-means-clustering-in-SAS-comparing-PROC-FASTCLUS-and-PROC/ta-p/221369

I was wondering if there was a way to derive the optimal seed for the FASTCLUS and HPCLUS procedures combined, as demonstrated in part 3 of the link.

Many thanks,
Ria.

Re: Initial seed for clustering

PGStats — Sun, 03 Jan 2016 19:49:09 GMT

I don't have access to EM. But beware of this: The word seed has different meanings for FASTCLUS and HPCLUS. For the former it is a set of cluster centroids. For the later it is a starting number for pseudo-random number generation.

Part 3 of the link proposes a way to select the optimal number of clusters (k) using HPCLUS and then to feed that number into FASTCLUS to get better clusters. In my view, it is only worth doing that for large problems. For smaller problems, running FASTCLUS reperatedly over a range of k-values does the job.