Re: Prediction in clustering

[Art Kendall]

There are methods in many packages.  That is the essential difference between classification and clustering.  For example, Since the mid-70's, SPSS's Discriminant Function Analysis procedure has had the ability to have some cases (rows) in the data file be "unclassified".  After it finds the discriminating functions, it applies the function to the unclassified cases and assigns them to the closest group, giving also the probability of membership of a case for each group and a probability of a case that is assigned to this group being so far from the centroid. Many of the other procedures in SPSS (e.g., Quick Cluster (k-means), TWOSTEP, TREE, REGRESSION, etc.) have explicit options to save a  model so that the model can be applied to cases that were not used in formulating the model. P.S. the TWOSTEP clustering procedure in SPSS gives an AIC and BIC for numbers of clusters in a requested range.  This can help you in determining the number to retain. Art [EMAIL PROTECTED] Social Research Consultants University Park, MD  USA (301) 864-5570 Jay Liu wrote: Dear all,   Apart from how to determine the number of clusters, another difficulty in clustering (I think) is how to predict cluster memberships of new data. This is very straight forward in classification but I can't think of a single clustering method I know can do this. I guess some model-based techniques maybe can do this but frankly, I have no clue at all.     Jay.