Clustering and scoring a new data

[Minnie]

I have done a lot of clustering, but I haven't had to score a new data file using the saved centroids. One thing I wonder is if we have all the variables for the clustering, why not running the cluster analysis with the new data instead of using the saved centroids. In the predictive model, we build a model based on the outcome we already know and estimate the probability for the unknown outcome. In the clustering, even in the initial classification, we still estimate the group membership, don't we?
I would appreciate any feedbacks.

[editor]

Clustering is a very "imprecise" science, because it is hard to say when
clusters are correct.

If you already have some meaningful clusters, then you may want to score the new data against the existing clusters. If the new data is comparable in size to old data, or if previous clusters have no meaning, then you can generate new clusters on the new data.

Gregory Piatetsky

[Minnie]

Thanks for your reply! Has anyone done this: generated clusters with a set of variables and then score a new data based on some of the original variables? We are collecting survey data and will generate clusters based on the survey items and demographic variable. Then I am going to score the entire customer database. I won't have the survey items in the customer database, but I would like to identify the entire customer clusters based on the partial data. Does this sound totally non-sense, or is this something people do? Thanks!