Partition Around Medoids (PAM) Clustering in SAS

[sasuser_621]

Apologies for the simple question. Is there a way to implement Partition Around Medoids (PAM) clustering, or k-medoids generally, in SAS?

I have looked at the official SAS documentation on clustering (also attached). There does not seem to be any procedure that uses k-medoids for clustering, unless I overlooked it.

If there is some SAS code to implement this, or to implement algorithms that apply PAM (e.g. CLARA from Kaufmann & Rousseeuw 1990, CLARANS from Ng & Han 1994), that would be a huge help.

Thank you!

[DougWielenga]

I have looked at the official SAS documentation on clustering (also attached). There does not seem to be any procedure that uses k-medoids for clustering, unless I overlooked it.

The k-medoids algorithm is similar to k-means but uses actual observations rather than 'center of mass' to identify the middle of each cluster.   Some authors suggest this makes the k-medoid approach more robust against outliers than k-means, but its requirement to compute all pairwise differences between each point in the cluster means that this particular approach does not scale well and is therefore note likely a good choice for most data mining problems. 

SAS Enterprise Miner uses observations which are sufficiently separated to create the initial cluster seeds before assigning the observations to clusters, but the final seeds are not required to be actual values.   I do not know of a way to constrain it to mimic the k-medoid approach but neither do I see that it would have a clear advantage in most cases over k-means which you can do with the FASTCLUS procedure or with the Cluster node which uses both FASTCLUS to get initial seeds and CLUSTER to group those seeds.

Hope this helps!

Doug