Replies to This Discussion

Permalink Reply by DataLLigence on September 8, 2008 at 11:57pm

hi tina,

i replied on the KDnuggets forum but i guess you are not checking that thread anymore....

1. you will find lots of documentation on deciding the number of clusters but as far as interpretation is concerned, it's pretty much subjective just like the interpretaion of a factor analysis result. a lot will depend on the specific business/domain for which you are analyzing the data, and the business requirement of your client.

2. you can use PROC CLUSTER. specify the variables you want with VAR, and if you want to use a non-Euclidean distance for clustering, you can compute a distance matrix using PROC DISTANCE. you'll find details on the SAS online help doc.

i also guess your dataset will have mixed variable types. there are a lot of opinions on which clustering method should be used for such data but i find SPSS's two-step cluster the best as far as i know.

rgds,

• ▶ Reply

Permalink Reply by Atul on September 10, 2008 at 8:01am

you can find a lot of documentation about this online. The following link gives detailed description about PROC CLUSTER in SAS.

http://www.otago.ac.nz/sas/stat/chap23/sect2.htm

Hope this helps.

• ▶ Reply

Permalink Reply by Ron Ritter on October 3, 2008 at 2:59am

I agree with Romakanta regarding the SPSS's two-steps cluster method to be the best as far as i know. I've done a segmentation project for one of Israel airlines and got good results using the two-steps method as it can handle both nominal and numerical variables. The lesson I've learned is to aware to the business goal and to strive answer it using the least amount of variables. If you would like i can describe the airline project in more details but i advise you to read about SPSS's two-steps cluster method first.
Regards,


Ron

• ▶ Reply

Permalink Reply by Eleazar on February 16, 2009 at 7:03pm

Hi, I'm also doing a research about consumer segmentation. In this case consumers buying products or service online.
I considered seven variables (after doing a factor analysis) and in conjoint with socio-economic variables, conform the dependent variables.

I used CHAID for the segmentation and I found four segments of customers each with their own decision rules.
Actually I'm developing an Agent Based Simulation to model online consumer behavior. Specifically creating agents representing the four segments founded.

• ▶ Reply

Permalink Reply by Matej Karpisek on February 19, 2009 at 4:00am

Hi Tina,

a small remark:
If you are looking for the group of companies with the highest potential to buy you should go for a propensity model...

regards,
Matej

• ▶ Reply

Permalink Reply by Vinayak NV on February 20, 2009 at 12:06am

Proc Varclus seems to be the best method for customer segmentation if you are analyzing data at product level.

Interpreting the tree is based on the connecting nodes (clusters). Which are all the products highly associated (within and b/w clusters).

Many Thanks,

Vinayak

• ▶ Reply

Permalink Reply by Isaac on February 21, 2009 at 12:41pm

I agree with all on using SPSS's 2-step procedure which works great for continuous and categorical data. I used it to Segment member bases for AOL and Allstate and the algorithms ran in under 15 minutes and interpretation was straight forward and passed face validity.

Regards,
Isaac Turk

• ▶ Reply

Permalink Reply by sheh on March 23, 2009 at 12:05pm

what does one do when there are almost 130 different product categories--example - i have SAS i see the objective is to cluster customers who buy similar products. what is the best way to deal with missing values?

customer id product1 product2 product3 ------ product130
111 1 . 1 .
123 . . . .
222 . . 1 .
..
999 . . . 1

• ▶ Reply