for what do we need training accuracy?

Fred12

hi,

 

I sometimes read about training accuracy, it should often be around 99%.. but my question is: Why does one need training accuracy? for what is it important, because usually, it gives no indication as how well your models expected performance is on new  test data... so why does one need that?

IngoRM

Hi,

 

See my recent post in this discussion here:

 

http://community.rapidminer.com/t5/RapidMiner-Studio/what-is-training-accuracy-testing-accuracy/m-p/32499#U32499

 

In short, you should NEVER focus on training error.  Any guidance like the 99% stuff is just not true (see my k-NN example in the other text).  The only applicable scenario is to detect overfitting but even here I personally do not find it too useful.

 

Cheers,

Ingo