Any RapidMiner experts out there want to help a n00b?

mohinipoojari16

I've been struggling with this for a few days and could use some gentle prodding in the right direction.

I have approx. 24k example rows, with 25 real (double's) attributes, and a nominal label. Each example represents a snapshot of scientific measurements at a moment in time (g-forces, magnometer, etc.) and the nominal label is essentially a boolean ("was the event happening?"). I'm trying to build a model (preferably a formula) that can predict the boolean output, or provide some sort of numerical "confidence".

Here's the issues I'm having:

Almost everything I do is running out of memory, I have 3GB of RAM devoted to the RapidMiner JVM.
In the event that I do get a model to self learn, I end up with something that has "97% accuracy", but always predicts one of the boolean values (e.g. it's 97% accurate and not 100% accurate, because it always predicts "false" and never predicts a "true")
I'm thinking some of my attributes are insignificant to the boolean result, but I don't understand how to identify which ones and eliminate them. I also think I'm wasting a lot of time trying each model type out (LibSVM, Neural Net, etc.) when the guru's would probably know which model applies to this type of data/problem.

Thanks for any help.

lionelderkrikor

Hi @mohinipoojari16,

 

 - It seems that you have imbalanced data. In this case, you can Sample your data : Your global accuracy will decrease, but 

your Recall (was the event happening? = True) will increase : your model will be able to predict some "true".

Take a look at this thread.

 

 - To perform "feature selection", RapidMiner propose many operators : 

Personnaly, I find Optimize Selection (Evolutionary) operator performant.

 

I hope it helps,

 

Regards,

 

Lionel