How to procede with Nominal Features in Artificial Neural Network

Hello friends:

I'm doing a predictive model for predicting a numeric variable and intend to use an 
artificial neural network. In KNIME, model variables must be between 0 and 1. 
As I have some categorical variables, some other nominal and ordinal, I would like 
to know what the best strategy to transform the data of the latter variables. 
I plan to use the One2Many Node, but some times, depending of classes of nominal feature,
the number of new features tends to grow a lot. The other way is to treat the categories as 
if they were numeric codes and then normalize [ 0,1]. 

Can anyone give me any advice on how I proceed?

Thank you

Gabriel

CHILE

 

Hi Gabriel,

this is a common problem when working with Machine Learning. Here is a short but effective article describing some options:

http://www.kdnuggets.com/2015/12/beyond-one-hot-exploration-categorical-variables.html

One-hot is what the One2Many node implements.

Another option not mentioned in the article above is using the so called "hashing trick":

https://en.wikipedia.org/wiki/Feature_hashing

Which one to choose depends largely on your data and available computational power.

Cheers,
Marco.

Dear Marco:

Thanks for your answer. I will read links you have shared with me. 

Gabriel.