How is data split into training / test sets in rapidminer GO?

cramsden

I am using rapid miner go and would like to know how data is being split into training / test sets and if its the same for each method (deep learning, gradient boosted trees etc.).

Is there anywhere in the docs that say first 70% of rows are used for training or something like this?

Thank you

alebo

Hi Chris, 
We use a 60/40 split for every model. If the target column is nominal, Go builds random subsets and ensures that value distribution is the same as in the original dataset. Otherwise, Go builds subsets randomly. 
Regards,
Andras

cramsden

Thank you,

Is there anymore information on this?  I am new to data science and self teaching, so I'm a bit confused by the terminology.  

I am asking because I am noticing a difference in the predictive power of my models based on which order the data set they were built on was originally uploaded.

To clarify, the data is 60 / 40 split but what goes into the 60 and 40 respectively is done randomly but ensuring the same distribution is kept?

Or is it the first 60% of rows and last 40% of rows for the split?

alebo

You can find free learning materials on https://academy.rapidminer.com/. I would recommend checking it out.
For example, on validation I found the following materials, that could help you:

• White paper: https://academy.rapidminer.com/learn/article/how-to-correctly-validate-machine-learning-models 
• Videos: https://academy.rapidminer.com/learn/video/validation-demo, https://academy.rapidminer.com/learn/video/validating-a-model

Unfortunately, these are mostly focused on RapidMiner Studio processes, but Go uses the underlying data science practices. You might want to experiment with Studio as well, as it gives you lot more flexibility than Go.

When splitting data, Go always shuffles the dataset. In case of a nominal (categorical) label, Go ensures the same distribution. 

Regards,
Andras