KBinfo measure differs with same training data

[jordimarch]

Hello,

I've got a problem with my RBF network. I thought I could use Kononenko & Bratko index (KBInfo) to get more accuracy on the final prediction. The problem is that KBInfo varies from run to run and I don't understand why.

For example, with a training file with 17000 instances I initially had a KBInfo of -0.89 and later on, when I had say 40 more instances in my training data, then KBInfo for the exact same individual test as before got a positive 0.3. OK, I though if I regenerated my training data with the original 17000 instances I would get the same -0.89 as the day before, but not, now, with the old training data I get a KBInfo of 0.29....

Haven't got the time to study in more detail what is happening. I'd like to know if someone has experienced this before and what can be causing this high variance in the KBInfo.

Thx & regards,
Jordi

[jordimarch]

By the way, what do you use as final prediction for a new "row"?

I'm using confusionMatrix plus the Kononenko & Bratko info to get the final accuracy, but maybe there are much better approaches.

Thx,
Jordi