Using multiple data type for neural networks prediction models

gabriel02v

Hi,

I need a little help regarding implementation of neural network  where the default model “Neural Net” only allows me numeric attributes as dependent variables, and no nominal attributes (either character, binomial or polynomial). My question is what kind of model fits better to my given situation: I want to build a model which it will contain multiple data types (binomial, polynomial, and numeric ).

Thank you

IngoRM

Hi,

 

You can use the new "Deep Learning" operator introduced in version 7.3.  This can work also on non-numerical attributes.

 

Alternatively, you can also transform your attributes into numerical attributes first with a couple of preprocessing operators, namely "Nominal to Binominal" followed by "Nominal to Numerical".

 

Hope that helps,

Ingo

IngoRM

Hi,

 

I think there is a parameter called "reproducible" or something similar which you need to activate.

 

Cheers,

Ingo

Telcontar120

@gabriel02v you may also be interested in this very helpful web app from RapidMiner, which lets you see which algorthims are suitable for different types of datasets, based on the data types of the label and the attributes you have.  The filtering options on the left are quite useful for seeing which approaches you might want to try without having to do additional transformations.

 

http://mod.rapidminer.com/#app

 

 

gabriel02v

Hi. thank you very much for advices. Finally I implemented "Deep Learning", and now in order to improve the prediction ratio or to improve the model itself I have to try different values of the parameters, and I have a problem in this sense, because each time I click run, I get different output result which it's not okay because I can't compare the versions of the model. I'm aware of the fact that Deep Leaning operator uses the so called random weight gradient descent initialisation function. What can I do in order to use the same random weights values every time I click run button? 

gabriel02v

It seems to be working. Now after I tried what variabiles have the best impact in my model, I want to to know which indicators should I look at in order to validate a deep learning model. I noticed from performance vector that my model has about 63% class prediction, which it's quite good and now here is my question: is R squared so important as it is in a linear regression model?, because as as I said the class prediction rate of 63% is satisfying me for the moment but the R squaread is less than 50%, only 31.45% more exactly.