Answered by:
BPM with EP for nominal / ordinal data
Question

Hi all,
I'm trying to use BPM and EP to predict if a future likelihood will occur (BPM will buy example). My dataset has about 25 variables (consisting of nominal and ordinal data) that will be used to train the BPM, however, within the 25 variables some are nominal data. What would be the best way of handling this type of data to try to get accurate results for the bernoulli likelihood?
Thursday, January 26, 2012 6:59 PM
Answers

You can represent categorical data as a 1 of N code. If the catagorical values take on large numbers of values you can use the sparse version of the BPM.
John
 Marked as answer by vba123 Monday, January 30, 2012 1:57 PM
Friday, January 27, 2012 9:44 AMOwner
All replies

You can represent categorical data as a 1 of N code. If the catagorical values take on large numbers of values you can use the sparse version of the BPM.
John
 Marked as answer by vba123 Monday, January 30, 2012 1:57 PM
Friday, January 27, 2012 9:44 AMOwner 
Thanks. That's what I figured regarding 1 of N.Monday, January 30, 2012 1:57 PM

How does BPM (willBuy example) generalize on the categorical 1 of N data? For example, does it look at each value discretely? Or does it average the values based on how close (numerically) they are coming to the values in the training set? I'm hoping it treats each value independent of an ordinal series.
Thursday, March 8, 2012 9:00 PM 
If you run to convergence, it should not matter what order the data appears in.
Friday, March 9, 2012 9:18 PMOwner