Clustering a large dataset with mixed variable types

[roopk]

I have a dataset that has 700,000 rows and various variables with mixed data-types: categorical, numeric and binary. I have read several suggestions on how to cluster categorical data but still couldn't find a solution for my problem. The common suggestions are listed below:

1) Use proc distance for the categorical variables to get a distance matrix and then use proc cluster: I can't do that as my dataset is too big to be handled by proc cluster. Another constraint is that even proc fastclus can handle a large dataset but it doesn't work with distance matrix or anything other than numeric data.

2) I don't find "creating dummy variables for the categorical variables" a good solution either as I feel the clusters would be hard to interpret.

3) Another suggestion was to use HPCLUS but HPCLUS can either use only categorical variables or only numeric interval variables to perform clustering but it does not perform clustering for mixed levels of input variables

Is there any other solution/proc available in SAS that could perform clustering on a large dataset with mixed variable types? 

[Reeza]

How many variables do you have? 700k is a lot but I would have assumed SAS could handle it. Can you change your memory settings to allow it to be processed? 

[roopk]

There are 100+ variables in total but I will do variable selection before using cluster analysis. I might start with only 8-10 variables but the issue is I don't know how to do cluster analysis with binary/categorical variables.

[PGStats · Accepted solution]

EM has HPCLUS which should handle larger problems.

[roopk]

Can it handle mixed variable types too?

[PGStats]

The doc says it does, check

http://support.sas.com/documentation/onlinedoc/miner/em14/emhpprcref.pdf

[roopk]

Thank you so much!

[brintonj]

Pg 82: "HPCLUS procedure uses only numeric interval variables or only nominal variables to perform clustering. It does not perform clustering for mixed levels of input variables."

[PGStats]

My bad. Out of the HP parallel processing procedures I only use HPSPLIT which allows mixed types. I wrongly assumed that HPCLUS would do the same.

Thanks @brintonj for setting the record strait.